THE ORIGINS OF ANTISOCIAL BEHAVIOR: INFLUENCES OF GENES, DEVELOPMENT, 
AND CONTEXT 

By 

Sarah Lynn Carroll 

A DISSERTATION 

Submitted to 
Michigan State University 
in partial fulfillment of the requirements   
for the degree of 

Psychology – Doctor of Philosophy 

2024 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ABSTRACT 

Developmental trajectories of youth antisocial behavior (ASB; i.e., physical aggression and rule 

violations) unfold via dynamic interplay between individual predispositions and the contexts in 

which their development is embedded. Indeed, prior work has implicated both genetic influences 

and the broader environmental context (i.e., family, peers, and neighborhoods) in the 

emergence of youth behavioral problems, both cross-sectionally and over time. Nevertheless, 

much remains unknown about the developmental origins of ASB. Extant longitudinal studies 

examining its genetic and environmental etiology have done so across relatively brief windows 

of development, and nearly all longitudinal phenotypic studies have relied on either person-

centered or variable-centered approaches but not both, with key downstream consequences for 

our understanding of the factors that give rise to the development of ASB. The present studies 

addressed these gaps in the literature by examining trajectories of ASB across early 

development via a series of behavioral genetic, variable-centered, and person-centered 

analyses, with familial, peer, and neighborhood characteristics considered as moderators. 

Participants were drawn from the Twin Study of Behavioral and Emotional Development in 

Children (TBED-C; N = 1,030 twin pairs) and MTwiNS projects within the Michigan State 

University Twin Registry. Both the TBED-C and its longitudinal follow-up, MTwiNS, were 

enriched for neighborhood disadvantage. Study 1 made use of a series of longitudinal growth 

curve models and classical twin analyses to examine the genetic and environmental origins of 

ASB from the preschool years through to emerging adulthood, with neighborhood disadvantage 

considered as a phenotypic predictor. Results indicated a mean-level decline in ASB that was 

shaped by both genetic influences and nonshared (i.e., person-specific) environmental factors. 

Exposure to neighborhood disadvantage predicted higher levels of ASB at baseline but a 

somewhat more rapid age-related decline. Studies 2 and 3 built on these findings by 

disambiguating physical aggression (AGG) and non-aggressive rule-breaking (RB), 

respectively, and by incorporating person-centered statistical methods as well as variable-

 
 
centered methods. In Study 2, the development of AGG was examined from middle childhood 

through to emerging adulthood via growth curve modeling and mixture modeling. Both sets of 

analyses identified age-related declines in AGG, but only the person-centered models 

elucidated the factors (i.e., low parent-child conflict and familial affluence) that interrupted 

trajectories of elevated AGG. Likewise, in Study 3, variable-centered analyses were integral to 

modeling the mean-level increase in RB for the full sample, whereas the person-centered 

models implicated peer delinquency as well as family and neighborhood socioeconomic 

disadvantage as predictors of persistent and escalating RB trajectories. Altogether, these 

findings underscore the non-interchangeable contributions of the family, neighborhood, and 

peer contexts to trajectories of ASB across nearly the entire early developmental period. 

Moreover, integration of findings from variable-centered and person-centered statistical 

approaches demonstrated the potential to illuminate risk factors predicting persistent behavioral 

problems and protective factors predicting recovery, which may have important methodological 

implications for future studies of developmental psychopathology. 

 
 
 
 
 
 
 
 
 
 
 
ACKNOWLEDGEMENTS 

This dissertation would not have been possible without Dr. Alex Burt’s guidance, 

expertise, and enthusiasm for twin research and Dr. Kelly Klump’s, Dr. Alytia Levendosky’s, and 

Dr. Amber Pearson’s insightful feedback as part of my committee. I am deeply grateful to Dr. 

Burt, my committee chair and advisor, for always challenging me to become a better researcher 

and for believing that it was, in fact, possible for me to integrate behavior genetic and 

developmental methods. I am also grateful for the support of all of the clinical faculty, 

particularly Dr. Levendosky, whose kind words of encouragement throughout grad school have 

always lifted me up. I would also like to thank my funding sources, including the National 

Institute of Child Health and Human Development and the Michigan State University Graduate 

School, as well as the research assistants and staff in the Burt Lab for their substantial time and 

effort in collecting data. 

My journey in clinical psychology began in Dr. Eric Turkheimer’s lab at the University of 

Virginia. I am thankful for Dr. Turkheimer’s continuous support since my first days as a research 

assistant, and particularly for his generosity in helping me with ridiculously complicated 

statistical analyses every spring. Getting to see him at the Behavior Genetics Association’s 

meeting each year is always a conference highlight. I am also immeasurably grateful to Erinn 

Falls and Stephanie Rudlaff, whose support has made such a difference in my life both 

personally and professionally. 

I am also thankful for my parents, brother, and friends, who have supported me in my 

work toward becoming a clinical psychologist. I couldn’t have imagined a better friend than 

Megan Mikhail to walk with me through grad school, and I am so very excited to watch (the 

future) Dr. Mikhail’s career continue to unfold. In addition, Jack Baker has been a continuous 

source of support and affirmation, particularly in my clinical training; he is one of several 

clinicians who I admire and seek to emulate. I am also thankful for Maddy Baker, who has been 

a dear friend for more than 20 years and will be for the rest of my life, and Ben Vogel, whose 

iv 

 
humor and warmth are such a delight. Throughout grad school, I have also fondly remembered 

my wonderful grandmother, Ruby, whose nursing career and deep compassion for others 

continue to inspire me. 

I have found so much meaning and joy in being part of the Eastminster Presbyterian 

Church community over the past six years. Thank you to the members of Eastminster for 

welcoming, encouraging, and loving me since my first week in East Lansing. To the extent that 

I’ve been able to embody that same welcoming spirit in the therapy room, I’m a better clinician 

for having known the people there. I am also grateful to the clients with whom I’ve gotten to 

work, some on a long-term basis and others briefly. Being able to walk with them for any length 

of time in their efforts to become happier, healthier, and more self-confident has been a privilege 

and a joy. I am grateful to my clinical supervisors, particularly Dr. Natalie Moser, Dr. Jason 

Stentoumis, and Dr. Carol Laub, for bolstering my confidence and providing an outstanding 

example of what ethical, compassionate mental healthcare looks like.    

Lastly, I am thankful for my precious rescue cats, Francie and Laurie, who have been a 

continuous source of cuddles and comic relief. 

v 

 
 
 
 
 
 
 
 
 
 
 
 
TABLE OF CONTENTS 

GENERAL INTRODUCTION…………………………………………………………………………...1 

CHAPTER 1: CONTINUITY AND CHANGE IN THE GENETIC AND ENVIRONMENTAL 
ETIOLOGY OF YOUTH ANTISOCIAL BEHAVIOR………………………………………………….9 

CHAPTER 2: UNDERSTANDING DESISTANCE FROM AGGRESSION: AN EMPIRICAL 
INTEGRATION OF PERSON-CENTERED AND VARIABLE-CENTERED APPROACHES…..35                              

CHAPTER 3: UNDERSTANDING HETEROGENEITY IN TRAJECTORIES OF RULE-
BREAKING VIA INTEGRATION OF VARIABLE-CENTERED AND PERSON-CENTERED 
METHODS……………………………………………………………………………………………...67 

GENERAL DISCUSSION………………………………………………………………………….....99 

BIBLIOGRAPHY……………………………………………………………………………………...108 

APPENDIX A: SUPPLEMENTAL MATERIAL FOR CHAPTER 1……………………………….122 

APPENDIX B: SUPPLEMENTAL MATERIAL FOR CHAPTER 2……………………………….124 

APPENDIX C: SUPPLEMENTAL MATERIAL FOR CHAPTER 3……………………………….142 

vi 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
GENERAL INTRODUCTION 

Youth antisocial behavior (ASB) encompasses a broad range of behaviors that violate 

societal norms and/or the basic rights of others (American Psychiatric Association, 2013). 

Children who demonstrate high levels of behavior problems often do so throughout early 

development, culminating in a series of poor outcomes by adulthood (e.g., poor mental and 

physical health, financial difficulties, incarceration; Odgers et al., 2008). Despite this high level 

of rank-order stability, however, ASB is far from a static trait. Indeed, a large body of literature 

has found the absolute frequency of ASB and its specific behavioral manifestations to vary 

dramatically with age. Specifically, engagement in physical aggression (AGG; e.g., hitting, 

kicking) declines throughout childhood and adolescence for most youth, whereas non-

aggressive rule-breaking (RB; e.g., lying, stealing) typically peaks in frequency during 

adolescence (Burt, 2012; Moffitt, 1993, 2003). Moreover, there are pronounced individual 

differences in the development of ASB, with youth differing in their engagement at baseline as 

well as in the magnitude and direction of their change over time (e.g., Doornward et al., 2012; 

NICHD Early Child Care Research Network & Arsenio, 2004). 

Individual differences in the frequency and specific manifestation of ASB have been 

linked to numerous factors. For instance, biological sex consistently predicts engagement in 

ASB, with males demonstrating markedly higher rates of physical AGG, and somewhat higher 

rates of RB, across the lifespan relative to females (Moffitt, 2003). Differences in the frequency 

of ASB are also observed across families and across societies (Rescorla et al., 2007), reflecting 

genetic influences as well as the effects of the rearing environment (Burt, 2022). The 

neighborhood and peer contexts are also critical predictors, such that children exposed to 

concentrated poverty and/or delinquent peers exhibit far more behavioral problems on average 

than other youth (Leventhal & Brooks-Gunn, 2000; Mann et al., 2015). Findings from 

longitudinal studies suggest that discrepancies in ASB engagement across neighborhoods 

persist throughout adolescence and into early adulthood (e.g., Lacourse, Dupéré, & Loeber, 

1 

 
 
2008; Maldonado-Molina, Reingle, et al., 2010; Spano et al., 2010). 

Although extant longitudinal research has shed light on trajectories of ASB, much 

remains unknown about its developmental origins. For instance, a small body of research has 

leveraged both longitudinal and classical twin methods to examine the genetic and 

environmental origins of ASB (e.g., Eley et al., 2003; Burt, McGue, Carter, & Iacono, 2007), but 

most studies have done so across a relatively brief window of development (e.g., middle 

childhood through early adolescence). As a result, it is unclear how genetic predispositions and 

environmental risk/protective factors differentially affect stability and change in ASB across 

multiple developmental stages. The genetic factors contributing to early childhood ASB, for 

example, could be the same ones that underlie stability during late adolescence, which would 

indicate that the rank-order stability in ASB observed across early development is largely 

genetic in origin. Alternatively, and consistent with research identifying age-related increases in 

the heritability of ASB (Rhee & Waldman, 2002), novel genetic influences may come into play 

during adolescence and contribute to systematic change. It is also unclear whether 

environmental influences on ASB are largely transient and unsystematic, or whether these 

factors systematically influence behavioral trajectories. In short, because no genetically 

informative study to date has examined continuity and change in ASB across all of early 

development, it is unclear to what extent genetic and environmental influences, respectively, are 

continuous from early childhood through emerging adulthood. 

Extant phenotypic studies have also yielded a somewhat limited picture of development. 

Specifically, almost all longitudinal studies rely on only one of two statistical approaches: 

variable-centered or person-centered. Variable-centered analyses use growth curve models to 

quantify the mean change(s) over time in the outcome of interest across the entire sample 

(McArdle & Epstein, 1987; see Figure 1.0a), whereas person-centered approaches extract 

subgroups exhibiting distinct trajectories over time (Nagin, 1999; see Figure 1.0b). Each of 

these approaches necessarily provides a partial picture of development. In variable-centered 

2 

 
studies, for instance, the growth curve representing average change may not capture the 

heterogeneity observed in most samples. The primary strength of person-centered methods lies 

in their ability to identify subgroups of youth following trajectories of either persistent 

engagement or recovery, as well as elucidate the risk and protective factors, respectively, that 

distinguish these youth from their peers. Nevertheless, person-centered studies of ASB are 

critiqued for subdividing a continuous trait into qualitatively distinct groups, with some 

researchers arguing that ASB is best-represented using variable-centered models (Walters, 

2011; Walters & Ruscio, 2013). Given the unique strengths and limitations of each method, the 

use of both is likely to yield a far more complete picture of development than either could 

provide by itself. 

In sum, ASB is a complex trait, influenced by a myriad of genetic and environmental 

factors that may or may not be continuous over time. Moreover, the frequency of engagement 

varies considerably within persons over time as well as between persons at any given timepoint. 

To represent the development of ASB in a way that captures this complexity, longitudinal 

studies must leverage a wide range of statistical methods that disentangle genetic and 

environmental influences and model continuity and change not only across individuals but also 

across trajectory groups. The present studies sought to do just this by examining trajectories of 

ASB from childhood through to emerging adulthood via a series of variable-centered, person-

centered, and classical twin analyses. As detailed below, we first elucidated the genetic and 

environmental origins of ASB trajectories in general. We then examined the phenotypic 

development of AGG and RB, respectively, through two complementary sets of longitudinal 

analyses, which each yield distinct information about the ways in which family and 

neighborhood characteristics shape behavioral development. Thus, the studies included in this 

dissertation collectively aimed to advance our understanding of the dynamic developmental 

processes that give rise to youth externalizing trajectories.   

To accomplish this, the present studies made use of longitudinal data from the Twin 

3 

 
Study of Behavioral and Emotional Development in Children (TBED-C), a sample within the 

population-based Michigan State University Twin Registry and the only twin sample in the world 

specifically enriched for neighborhood disadvantage (MSUTR; Burt & Klump, 2019). Behavioral 

and emotional data relevant to the present studies were collected at up to five time points 

(between ages three and 22 years). Thus, this set of studies addressed the primary limitations 

of prior longitudinal work by examining engagement in ASB across multiple developmental 

stages and integrating results from two distinct statistical approaches. All studies examined the 

moderating role of neighborhood disadvantage on trajectories of continuity and trajectories of 

change in youth ASB.  

Overview of the three studies 

Study 1 examined the origins of stability and change in ASB from preschool through to 

emerging adulthood, with biological sex and neighborhood disadvantage considered as 

phenotypic predictors. Extant longitudinal twin studies have consistently found genetic 

influences to underlie rank-order stability in ASB, and nonshared environmental influences (e.g., 

differential parenting, peer groups) to contribute to change (e.g., Bartels et al., 2004; Burt et al., 

2007; Eley et al., 2003). The role of the shared environment (e.g., socioeconomic status, family 

mores) has appeared to be dependent on age, with studies in childhood samples reporting 

shared environmental influences to contribute to stability (e.g., van Beijsterveldt et al., 2003) 

while studies in older samples have found their effects to be negligible (e.g., Burt et al., 2007). 

However, as discussed above, these studies all assessed continuity and change in ASB over 

relatively brief time periods. Because no prior study has examined the etiology of ASB across all 

of early development, it is unclear to what extent genetic factors underlie stability in ASB from 

the preschool years through emerging adulthood, nor is it clear how shared and nonshared 

environmental factors differentially affect stability and change across multiple developmental 

stages. The present study thus made use of variable-centered multilevel growth curve modeling 

to examine the development of ASB from ages three to 22 years in participants from the TBED-

4 

 
C (N = 1,030 twin pairs) assessed up to five times. Biological sex and neighborhood 

disadvantage, defined via a composite index of area deprivation (see Table 1.0), were 

examined as predictors of the phenotypic intercept and slope. Classical twin modeling was then 

used to quantify the genetic, shared environmental, and nonshared environmental contributions 

to participants’ baseline levels of ASB (i.e., intercepts) and rates of change (i.e., slopes).  

Studies 2 and 3 built on Study 1 by 1) modeling behavioral trajectories via two distinct 

statistical approaches, and 2) distinguishing between physically aggressive and non-aggressive 

dimensions of ASB, respectively. As discussed earlier, variable-centered and person-centered 

approaches are rarely integrated, with no extant empirical study considering the ways in which 

their respective results may complement one another. The results from extant variable-centered 

and person-centered literatures were conceptually integrated in a recent review of longitudinal 

studies of ASB (Carroll, Mikhail, & Burt, 2023), but no empirical study to date has integrated the 

two approaches. Study 2 began to fill this gap in the literature by leveraging both approaches to 

examine desistance from AGG between childhood and emerging adulthood. We focused 

specifically on behavioral trajectories across middle childhood and adolescence, two key 

developmental periods characterized by significant changes in the frequency of engagement in 

ASB (Moffitt, 1993). Biological sex, family characteristics (i.e., household income, parent-child 

conflict, and parent-child nurturance), and neighborhood disadvantage were each evaluated as 

phenotypic moderators of AGG trajectories, with a focus on the protective factors predicting 

desistance. Study 3 leveraged the same modeling approaches to examine the development of 

RB, which follows a typical trajectory that is markedly different from that of AGG (Burt, 2012). 

Specifically, Study 3 sought to elucidate the risk factors predicting persistent and escalating 

engagement in RB, a pattern often observed during adolescence (e.g., Yan et al., 2021). 

Biological sex, household income, neighborhood disadvantage, and peer delinquency were 

examined as moderators of RB trajectories. 

Altogether, by conducting genetically informative analyses of ASB across multiple 

5 

 
developmental periods and integrating findings from two distinct statistical approaches, the 

present studies substantially advance our understanding of the patterns of desistance, 

persistence, and escalation that characterize the development of ASB across time and persons. 

Moreover, the demonstrated utility of variable-centered and person-centered approaches 

together may have important implications for future work examining trajectories of 

psychopathology. 

6 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Tables 

Table 1.0. Measures included in Area Deprivation Index 

Measure 

1.  Percent of population aged 25 and older with <9 years of education 
2.  Percent of population aged 25 and older with at least a high school 

diploma 

3.  Percent of population aged 16 and older in white-collar occupations 
4.  Median family income 
5.  Income disparity (ratio of households with <$10,000 income to households 

with ≥$50,000 income) 

6.  Median home value 
7.  Median gross rent 
8.  Median monthly mortgage 
9.  Percent of housing units owned by occupiers 
10. Percent of population aged 16 and older who are unemployed 
11. Percent of families below poverty level 
12. Percent of population below 150% of the poverty threshold 
13. Percent of households with children under age 18 headed by a single 

parent 

14. Percent of households without a motor vehicle 
15. Percent of households without a telephone 
16. Percent of occupied housing units without complete plumbing 
17. Percent of households with more than 1 person per room 

Note: Participating families’ ADI scores were determined by the level of deprivation in their 

Census tract (Study 1) and block group (Studies 2 and 3) based on all indices listed above. For 

additional details, see Singh (2003) and Kind & Buckingham (2018). 

7 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Figures 

Variable-centered: growth curve model

5

6

7

8

Age

9

10

11

Participant 1

Participant 2

Participant 3

Participant 4

Sample mean

Participant 5

Person-centered: mixture model

B
S
A

4
3.5
3
2.5
2
1.5
1
0.5
0

B
S
A

4
3.5
3
2.5
2
1.5
1
0.5
0

a.

5

6

7

8

Age

9

10

11

Group 1: low, stable

Group 2: moderate, desisting

Group 3: high, increasing

b.
Figure 1.0. Variable-centered (a) and person-centered (b) approaches to modeling the 

development of ASB across middle childhood in a hypothetical sample. The variable-centered 

approach identified a mean-level trend across the entire sample, whereas the person-centered 

analysis extracted three distinct trajectory groups. 

8 

 
 
 
 
 
 
 
 
 
 
CHAPTER 1: CONTINUITY AND CHANGE IN THE GENETIC AND ENVIRONMENTAL 

ETIOLOGY OF YOUTH ANTISOCIAL BEHAVIOR  

Abstract 

Background: Trajectories of youth antisocial behavior (ASB) are characterized by both 

continuity and change. Twin studies have further indicated that genetic factors underlie 

continuity, while environmental exposures unique to each child in a given family underlie 

change. However, most behavioral genetic studies have examined continuity and change during 

relatively brief windows of development (e.g., during childhood but not into adolescence). It is 

unclear whether these findings would persist when ASB trajectories are examined across 

multiple stages of early development (i.e., from early childhood into emerging adulthood). 

Methods: Our study sought to fill this gap by examining participants assessed up to five times 

between the ages of 3 and 22 years using an accelerated longitudinal design in the Michigan 

State University Twin Registry (MSUTR). We specifically examined the etiologies of stability and 

change via growth curve modeling and a series of univariate and bivariate twin analyses. 

Results: While participants exhibited moderate-to-high rank-order stability, mean levels of ASB 

decreased linearly with age. Genetic and nonshared environmental influences that were present 

in early childhood also contributed to both stability and change across development, while 

shared environmental contributions were negligible. In addition, genetic and nonshared 

environmental influences that were not yet present at the initial assessment contributed to 

change over time. Conclusions: Although ASB tended to decrease in frequency with age, 

participants who engaged in high levels of ASB during childhood generally continued to do so 

throughout development. Moreover, the genetic and nonshared environmental contributions to 

ASB early in development also shaped the magnitude of the decrease with age.  

9 

 
 
 
 
Introduction 

Childhood antisocial behavior (ASB) predicts a myriad of poor outcomes in adolescence 

and young adulthood, such as substance use, poor physical health, and internalizing pathology, 

as well as continued engagement in delinquent activities (e.g., Odgers et al., 2008). In children, 

ASB is characterized by persistent aggression, deceitfulness, property destruction, and/or rule 

violations (American Psychiatric Association, 2013). One of the defining features of ASB is its 

relatively high level of rank-order stability, with the same individuals typically exhibiting the 

highest levels of delinquent behavior across development. Despite this stability, mean levels of 

ASB decline throughout the first twenty or so years of life (e.g., Monahan et al., 2009), and there 

is considerable individual variation in the magnitude of this decline (Burt, 2012; Martino et al., 

2008). In short, extant research has clearly indicated that youth ASB trajectories are 

characterized by both stability and change, and that these patterns vary from person to person.  

To date, however, it is less clear what etiologic mechanisms underlie individual 

differences in these patterns of continuity and change. There are several competing a priori 

possibilities. The relatively high rank-order stability of ASB could be due to continuity in 

underlying genetic influences, while change over time could stem from specific environmental 

exposures (e.g., environmental risk factors could predict escalating behavior problems while 

environmental protective factors predict desistance). Alternatively, genetic contributions could 

change over time as different genes become (de)activated, while environmental factors could 

contribute to stability. One method for evaluating these competing hypotheses is the twin 

design, which compares monozygotic (MZ) and dizygotic (DZ) twins to disambiguate the genetic 

and environmental contributions to a given phenotype. By examining these contributions across 

multiple timepoints, twin researchers can clarify the origins of continuity and change.  

Prior longitudinal twin studies have begun to evaluate these possibilities. These studies 

have consistently implicated genetic factors as a major source of rank-order stability in ASB 

(Bartels et al., 2004; Burt et al., 2007; Eley et al., 2003; Lacourse et al., 2014; Pingault et al., 

10 

 
2015; Porsch et al., 2016; van Beijsterveldt et al., 2003). As an example, Burt et al. (2007) used 

latent growth curve modeling to examine the etiologic trajectory of ASB from late adolescence 

through early adulthood (approximately ages 17-25) in 626 twin pairs from the Minnesota Twin 

Family Study and found the same genetic factors to be present over time. These factors 

explained a moderate-to-large proportion of the variance in ASB at each timepoint and were 

largely responsible for trait stability. Nonshared environmental influences (or experiences that 

serve to differentiate children raised in the same family; e.g., peer groups) were found to 

underlie change over time. Shared environmental influences (experiences common to children 

raised in the same family; e.g., similar parenting) did not contribute to ASB at baseline or over 

time, consistent with research indicating that shared environmental influences on ASB become 

less salient (and nonshared more salient) with age (e.g., Tuvblad et al., 2011). Child and 

adolescent twin studies using liability threshold analyses, simplex modeling, or Cholesky 

decomposition modeling have reported somewhat similar results, finding that nonshared 

environmental factors largely contribute to change over time, while genetic influences contribute 

to stability (Bartels et al., 2004; Eley et al., 2003; van Beijsterveldt et al., 2003). However, these 

studies have also found evidence that shared environmental influences contribute to stability 

over time. In other words, shared environmental influences have been found to impact ASB 

development in younger samples, but their effects appear to be negligible by late adolescence.   

 Despite these consistencies in results, several questions remain unanswered. First, it is 

unclear whether the genetic factors contributing to adolescent and young adult ASB are the 

same as those contributing to child ASB. Although the studies discussed above found largely 

continuous genetic effects, most assessed continuity and change during a relatively brief 

window of development: early childhood through pre-adolescence, or middle childhood through 

early adolescence, or late adolescence through early adulthood. The most comprehensive 

etiologic study of ASB development (Pingault et al., 2015) began in early childhood (twins were 

assessed beginning at age 4 through 16 years) but did not assess participants in late 

11 

 
adolescence or emerging adulthood, two key developmental periods in the transition to adult 

social and occupational roles (Alink & Egeland, 2013). Indeed, no study to date has examined 

the etiologies of continuity and change in ASB across all of early development (i.e., the first 

twenty or so years of life). As such, we do not know to what extent genetic factors underlie 

stability in ASB from the preschool years through emerging adulthood, nor is it clear how shared 

and nonshared environmental factors differentially affect stability and change throughout this 

period.  

The latter uncertainty is particularly important, given that the shared environment has 

been identified as an important etiologic source of stability in ASB during childhood and early 

adolescence (Bartels et al., 2004; Eley et al., 2003; van Beijsterveldt et al., 2003), but does not 

appear to affect the trajectory of ASB during late adolescence (Burt et al., 2007). By contrast, 

nonshared environmental influences appear to be transient and idiosyncratic prior to adulthood, 

with nonshared environmental correlations for positive and negative affect and interpersonal 

warmth each decreasing monotonically in a matter of minutes or days (Burt et al., 2015). At 

some point during adolescence, however, these influences appear to become more enduring. 

Although a few studies have reported this increased stability of the non-shared environment as 

early as age 7 (e.g., van Beijsterveldt et al., 2003), most studies place it sometime in late 

adolescence (e.g., Hopwood et al., 2011). As such, results from studies of children identifying 

the nonshared environment as a source of change may not persist to older samples. In short, 

much is unknown about the developmental origins of ASB because no study of its etiologic 

trajectory has spanned early childhood through emerging adulthood. 

The aim of the present study was to address these gaps in the literature by examining 

the origins of stability and change in youth ASB from preschool through to emerging adulthood. 

We used up to five waves of data from the Michigan State University Twin Registry (MSUTR; 

Burt & Klump, 2019) collected across ages 3 to 22 years using an accelerated longitudinal 

design. We applied multilevel growth curve modeling, in which measurements were nested 

12 

 
within participants, to estimate participants’ baseline levels of ASB (i.e., intercepts) and rates of 

change with age (i.e., slopes). We subsequently used classical twin modeling to quantify the 

genetic, shared environmental, and nonshared environmental influences on the intercept and 

slope, respectively. Based on prior research, we hypothesized that genetic factors would 

underlie stability and nonshared environmental factors would underlie change. Given that 

shared environmental effects have been found to decrease with age, we did not expect them to 

significantly impact participants’ trajectories into emerging adulthood. 

Methods 

Participants 

Participants were drawn from the Twin Study of Behavioral and Emotional Development 

in Children (TBED-C), a sample within the population-based Michigan State University Twin 

Registry (MSUTR; Burt & Klump, 2019). The TBED-C includes both a population-based 

subsample (n=528 families) and an independent ‘at risk’ subsample of twin families residing in 

impoverished Census tracts (n=502 families). When combined, the overall sample thus 

comprised 1,030 twin pairs: 224 MZ male pairs, 211 DZ male pairs, 202 MZ female pairs, 207 

DZ female pairs, and 186 DZ opposite-sex pairs. Mean household income at the middle 

childhood assessment was $76,329 (SD=$45,650) in the population-based sample and $55,652 

(SD=$31,088) in the at-risk sample. Other recruitment details are reported at length in prior 

publications (e.g., Burt & Klump, 2019). Families across the two samples collectively identified 

as White (non-Latinx): 81%, Black: 10%, Latinx: 1%, Asian: 1%, Indigenous: 1%, and 

multiracial: 6%. These proportions are largely consistent with those for the population of the 

State of Michigan, based on data from the U.S. Census Bureau (http://www.Census.gov/) (e.g., 

White: 79%, Black: 14%). For all studies, parents provided informed consent, children provided 

informed assent, and families were compensated for their time.   

 Behavioral and emotional data relevant to the current study were collected at as many 

as five time points. All 1,030 twin families were assessed once in middle childhood (ages 6-11) 

13 

 
as part of the TBED-C. Those TBED-C twins residing in modestly-to-severely disadvantaged 

neighborhoods are currently being reassessed in-person as adolescents up to two times, 18-

months apart, through the Michigan Twin Neurogenetics Study (MTwiNS). The first of the 

MTwiNS assessments was conducted approximately 4-6 years after participation in TBED-C 

(ages 7-19; currently available for 354 families), while the second adolescent assessment was 

5-7 years after participation in TBED-C (ages 10-19; currently available for 188 families). TBED-

C families with twins between ages 11 and 22 were also recently recruited for an online 

assessment of youth psychopathology (N=637 families completed the online assessment). 

Finally, we were also able to link to data collected on TBED-C families as part of the population-

based Michigan Twins Project (MTP), an ongoing study of approximately 12,000 Michigan-born 

child and adolescent twin pairs (93.3% of TBED-C families were recruited out of the MTP). See 

Table 1.1 for additional details about the sample at each age, including sample sizes at each 

assessment.  

Notably, the MTP assessments were not completed in the same order or at the same 

ages across participating TBED-C families. For example, while most families were first 

assessed as part of the MTP (73.9%), others were first assessed as part of TBED-C and were 

assessed only later as part of the MTP (26.1%). Follow-up MTP assessments are also on-going. 

A portion of TBED-C families (N = 637; 61.8%; 56 of these were assessed twice) were re-

assessed approximately 5-8 years after their original participation in the MTP. To account for 

these irregularities in the ordering of data collection across the TBED-C/MTwiNS and the MTP, 

data were organized chronologically by age for each participating twin. As such, each 

assessment wave in the current study includes MTP and either TBED-C or MTwiNS 

assessments (see Table 1.2). A total of 677 pairs (66% of the sample) completed at least three 

of the assessments, while 96% (N=989) completed at least two. The highest number of 

assessments completed by any single participant was five. 

14 

 
 
Recruitment 

Because birth records are confidential in Michigan, we collaborated with the Michigan 

Department of Health and Human Services (MDHHS; formerly known as the Michigan 

Department of Community Health) to recruit families for all MSUTR twin studies (including the 

MTP and all waves of the TBED-C/MTwiNS). The MDHHS is the agency in charge of all vital 

records in the State of Michigan and thus has direct access to individual SSNs, full names, and 

birth dates. The MDHHS identifies twin pairs residing in Lower Michigan who meet age criteria 

for a given study and whose addresses or parents’ addresses (for twins who are minors) can be 

located either using driver’s license information obtained from the State of Michigan or the 

proprietary search engine used by police. Twins indicating interest in participation via pre-

stamped postcards or e-mails/calls to the MSUTR project office are then contacted by study 

staff to determine study eligibility and to schedule their assessments.  

Four recruitment mailings were used to ensure optimal twin participation. Overall, 

response rates across studies (56-85%) are on par with or better than those of other twin 

registries that use similar types of anonymous recruitment mailings and have thus far yielded 

largely representative samples. Families of the naturally-conceived twins in the large-scale 

MTP, for example, closely resemble families across the State of Michigan (Burt & Klump, 2013). 

The proportion of MTP families that identify as White, non-Hispanic (81.0%) is very similar to the 

80.2% indicated in state-wide Census data. Mean family incomes are also quite comparable 

($75,940 in the MTP versus $73,373 in the Census), as are the proportion of families with 

graduate or professional degrees (10.3% in the MTP versus 9.6% in the Census).   

Because 90+% of TBED-C families were recruited out of the MTP, we were able to use 

the MTP data to compare families who chose to participate in TBED-C with those who were 

recruited but did not participate. TBED-C families were generally representative of recruited but 

non-participating families. As compared to non-participating twins, participating twins reported 

similar levels of conduct problems, emotional symptoms, and hyperactivity (d ranged from -.08 

15 

 
to .01 in the population-based sample and .01 to .09 in the at-risk sample; all ns). Participating 

families also did not differ from non-participating families in paternal felony convictions (d = -.01 

and .13 for the population-based and the at-risk samples, respectively), rate of single parent 

homes (d = .10 and -.01 for the population-based and the at-risk samples, respectively), 

paternal years of education (both d ≤ .12), or maternal and paternal alcohol problems (d ranged 

from .03 to .05 across the two samples). However, participating mothers in both samples 

reported slightly more years of education (d = .17 and .26, both p < .05) than non-participating 

mothers. Maternal felony convictions differed across participating and non-participating families 

in the population-based sample (d = -.20; p < .05) but not in the at-risk sample (d = .02). In 

short, our recruitment procedures thus appear to yield samples that are representative of both 

recruited families and the general population of the State of Michigan. 

Procedure 

Some of the assessments, specifically the MTP assessments and the “online” 

assessment, were completed remotely by the twins’ primary caregiver, nearly always their 

mother. The TBED-C and MTwiNS assessments were completed in-person. For TBED-C, twins 

were assessed either at our East Lansing-based laboratories or at the family’s home. 

Questionnaires did not vary across the laboratory-based and family home assessments. For 

MTwiNS, the twins and their parent(s) completed an in-person assessment lasting 4-8 hours at 

either the East Lansing or Ann Arbor-based laboratories. 

Measures 

Youth Antisocial Behavior 

Youth ASB was assessed via maternal report at all ages. At the MTP assessments, 

participating twins’ mothers completed the Strengths and Difficulties Questionnaire (SDQ; 

Goodman, 2001), a 25-item measure in which parents rate the extent to which a series of 

statements describe the child’s behavior over the past six months using a three-point scale 

(0=not true to 2=certainly true). For these analyses, we focused on the Conduct Problems 

16 

 
subscale (5 items: hot temper, obedient (reverse-scored), fights, lies or cheats, steals; α=.60, 

.63, and .66 at MTP assessments 1, 2, and 3, respectively). Psychometric studies have found 

the SDQ to have satisfactory test-retest reliability (r>.85 for the Conduct Problems subscale) 

and to be highly correlated with other parent-report measures, including the Child Behavior 

Checklist (CBCL) (e.g., Muris et al., 2003). In addition, studies in samples spanning childhood 

and adolescence support the use of the parent-report SDQ as a screening measure that 

adequately distinguishes between community and clinical populations across age groups 

(Becker et al., 2004; He et al., 2013). 

At the in-person TBED-C and MTwiNS assessments, the twins’ mothers completed the 

CBCL (Achenbach & Rescorla, 2001), rating the extent to which a series of statements 

described the child’s behavior during the past six months on a three-point scale (0=never to 

2=often/mostly true). To maximize comparability with the SDQ, we constructed a scale using 5 

items on the CBCL that were analogous to those on the SDQ: hot temper, disobedient at home, 

fights, lies or cheats, steals from home (α=.66 in the TBED-C and .69 and .60, respectively, at 

the two MTwiNS assessments). While these items screen for behaviors that may manifest 

differently at different ages (e.g., lying, hot temper), all are relatively common across early 

development. As such, we believe that they adequately assess ASB across the broad age 

range included in our sample.  

Zygosity 

Zygosity was established using physical similarity questionnaires administered to the 

twins’ primary caregiver (Peeters et al., 1998). On average, the physical similarity 

questionnaires used by the MSUTR have accuracy rates of at least 95% when compared to 

DNA. 

Disadvantage 

Socioeconomic disadvantage was assessed using the Area Deprivation Index (ADI), a 

composite measure comprising 17 indices of Census-tract disadvantage (e.g., poverty rate, 

17 

 
income disparity). We recreated Kind & Buckingham’s index of disadvantage in our sample, as 

assessed via Census data collected from 2008 to 2012. The measures were weighted 

according to the factor loadings identified by Kind & Buckingham (2018), and weighted variables 

were summed to create a deprivation index score for each Census tract. Families were 

assigned a percentile indicating the level of deprivation in their Census tract relative to that of all 

Census tracts in Michigan. The mean ADI was 42.51 (SD=26.17) and ranged from 1 to 100.  

Data Analyses 

Phenotypic Analyses 

All analyses were conducted using Mplus 8.0 (Muthén & Muthén, 1998-2019). To 

examine phenotypic changes in ASB over time, we used a three-level growth curve model in 

which occasions of measurement (Level 1) were nested within participants (Level 2) who were 

nested within families (Level 3). These models capture both the average rate of change in ASB 

over the course of the study, as well as individual variability in change via random intercept and 

slope terms. Age was used as the index of time in these models and was centered at three 

years old, the youngest age in our sample. The intercept can thus be interpreted as the level of 

ASB at age 3.  

We initially estimated an unconditional growth model with a random linear slope, which 

allowed for interindividual variation in the rate of change over time. Models with non-linear 

slopes encountered serious convergence difficulties. Moreover, prior research consistently 

indicates that mean levels of ASB decline steadily throughout development (e.g., Monahan et 

al., 2009). We subsequently estimated a conditional growth model. In this model, questionnaire 

type (SDQ or CBCL) was added as a time-varying covariate on level 1 with a random slope. We 

included biological sex, ethnicity, and ADI as covariates of the random intercept and slope on 

level 2.  

Full-information maximum likelihood estimation was used to account for missing data, as 

prior simulations have shown it to be robust to at least 50% missing data (Enders & Bandalos, 

18 

 
2001). Moreover, accelerated longitudinal designs such as ours, which have planned missing 

data, have been found to have robust power despite small sample sizes at certain ages 

(Rhemtulla & Hancock, 2016). Data were log-transformed prior to analysis to better approximate 

normality. Random intercept and slope factor scores were generated for subsequent biometric 

analyses using maximum a posteriori (MAP) scoring (MacCallum, 2009). The individual factor 

scores were obtained from an unconditional, two-level model in which measurement occasions 

were nested within participants, as biometric twin models account for clustering within families. 

Biological sex, ethnicity, and ADI were regressed out of the factor scores prior to running the 

twin analyses (McGue & Bouchard, 1984).  

Biometric Twin Analyses 

A series of biometric twin models were then run to estimate the relative genetic and 

environmental influences on variability in the estimated intercept and slope factor scores. We 

used these factor score estimates rather than running a full biometric latent growth curve model 

due to the large variation in participant age at each timepoint (time is typically based on 

assessment schedule in biometric latent growth curve models). Time can be easily modelled via 

participants’ chronological age at a given time point in multilevel growth models, however, 

making it better-suited for study designs that involve an uneven schedule of assessments (Hox 

& Stoel, 2014), such as this one. Although each analytic framework has its practical 

advantages, structural equation and multilevel growth models are conceptually analogous.  

Classical twin models leverage the difference in the proportion of segregating genes 

shared between identical (MZ) twins, who share 100% of their genes, and fraternal (DZ) twins, 

who share an average of 50% of their segregating genes to estimate the relative contributions of 

genetic and environmental influences to the variance within observed behaviors (phenotypes). 

Phenotypic variance is decomposed into three variance components: additive genetic (A), 

shared environmental (C), and nonshared environmental (E). More information on twin modeling 

is provided elsewhere (Neale & Cardon, 1992). For the present study, we first computed the A, 

19 

 
C, and E estimates for the random intercept and slope factors, respectively. We then used a 

bivariate twin model to clarify the extent to which the etiologies of the slope and the intercept 

overlapped. 

Descriptive Statistics 

Results 

Descriptive statistics by age are shown in Table 1.1, while Table 1.2 contains 

correlations across assessment waves (operationalized here in person-specific chronological 

order because of the irregularities in when specific assessments were administered). 

Participants evidenced moderate-to-high rank-order stability in their reported ASB over time, 

with correlations ranging from .20 to .56. Paired-sample t tests further indicated that, within 

persons, ASB decreased significantly from the first assessment to the second (t(1913) = -7.97, 

p<.001), and from the second assessment to the third (t(1333) = -9.55, p<.001). Changes 

across subsequent assessments were not significant. In addition, mean ASB scores decreased 

steadily with age (see Figure 1.1). Collectively, these findings indicate that, while participants 

who exhibited high levels of ASB during childhood largely continued to do so during 

adolescence, the absolute level of ASB decreased significantly over time. Although not shown in 

Table 1.2, males had slightly higher ASB scores than females at the first two timepoints 

(Cohen’s d = .21 and .22, respectively; both p < .001), whereas there were no significant sex 

differences at assessments three, four, or five. Lastly, ADI was significantly correlated with ASB 

at the first two timepoints (r = .17 and .11, respectively, both p < .001), but not at the last three. 

Multilevel Modeling 

Phenotypic Results 

 In the baseline unconditional growth model, ASB decreased significantly over time 

(slope mean = -.03, p<.001), although there was significant interindividual variation in the 

magnitude of the decline (slope variance=.001, p<.001). The covariance between the intercept 

and the slope was negative, meaning that participants with higher ASB scores at baseline 

20 

 
tended to exhibit a more rapid decline in ASB over time. We next fitted a conditional growth 

model. Results are shown in Table 1.3. Sex significantly predicted both intercept and slope, 

such that male participants had higher ASB scores at baseline and displayed a more rapid 

decline with age. ADI was also a significant predictor of both the intercept and slope, with 

participants from more impoverished neighborhoods exhibiting higher levels of ASB at baseline 

and declining more rapidly over time. By contrast, race (a socially constructed category coded 

as white/non-white given the composition of our sample) did not significantly predict the 

intercept or the slope. 

Twin Model Results 

Univariate ACE models were first run using intercept and slope factor scores from the 

unconditional growth model, in order to estimate the genetic, shared environmental, and 

nonshared environmental contributions to stability and change in ASB. Standardized univariate 

variance estimates are presented in Table 1.4. For both the intercept and the slope, there were 

significant genetic and nonshared environmental contributions, but no significant shared 

environmental influences, although the estimated magnitude of the shared environmental 

variance for the intercept was non-zero (.10).  

The bivariate ACE model indicated both significant unique and overlapping genetic and 

nonshared environmental influences on the intercept and the slope factors. Path estimates are 

shown in Figure 1.2. More than one-third (38%) of the genetic variance in the slope factor was 

shared with the intercept. Thus, the genetic etiology of ASB development was due to both 

genetic influences that had emerged during the preschool years, as well as to novel genetic 

influences emerging later in development. Interestingly, one-third of the nonshared 

environmental variance in the slope was also present at baseline, indicating that our estimates 

of E did not represent solely transient, time-specific influences, but rather exhibited a fair 

amount of stability across development. Consistent with the univariate results, shared 

environmental contributions were not significant. Taken together, the genetic and nonshared 

21 

 
environmental influences on ASB in early childhood also appear to contribute to its stability 

across development. 

Discussion 

The aim of the present study was to elucidate genetic and environmental contributions to 

continuity and change in ASB from early childhood into emerging adulthood. To do so, we 

obtained estimates of participants’ baseline level of ASB and change over time via multilevel 

growth curve modeling. We then made use of a series of classical twin models to illuminate 

genetic, shared environmental, and nonshared environmental contributions to the intercept and 

the slope of ASB, as estimated via factor scores generated in the prior analyses. The results 

indicate that initial levels and change over time in ASB were due to both genetic and nonshared 

environmental influences, some of which overlapped. Neither initial level of ASB nor change 

over time were subject to significant shared environmental influences.  

Genetic influences were found to make important contributions to ASB in early life, as 

well as to change in ASB across development. Furthermore, more than one-third of the genetic 

contributions to change over time were already present at baseline (i.e., during the preschool 

years), indicating a fair amount of continuity in genetic influences throughout early development. 

These findings are consistent with those of other studies that have found prominent genetic 

influences on continuity in youth ASB, albeit during much shorter windows of development 

(Bartels et al., 2004; Burt et al., 2007; Eley et al., 2003; Porsch et al., 2016) or over longer 

periods that did not include late adolescence/emerging adulthood (Pingault et al., 2015). Our 

study extends these findings by indicating that genetic influences contribute to continuity in ASB 

across all of early development (i.e., the first 20 or so years of life). Genetic influences were also 

found to underlie change in ASB, with nearly two-thirds of the genetic variance in the slope 

representing novel influences that were not present at baseline. The emergence of novel 

genetic influences over the course of our study is unsurprising, given the broad age range (3 to 

22 years) represented in the sample. In addition, this pattern of results is consistent with those 

22 

 
of other studies of children and adolescents that have found genetic influences to contribute to 

both continuity and change in ASB (Bartels et al., 2004), particularly for nonaggressive rule-

breaking (Eley et al., 2003).  

Nevertheless, neither continuity nor change in ASB were due solely to genetic 

influences. Nonshared environmental variance played a considerable role in continuity across 

development in our sample, with fully one-third of the nonshared environmental contributions to 

the slope already present at baseline. Such findings stand in contrast to those of prior 

longitudinal studies of youth ASB, which have found the nonshared environment to exert largely 

transient effects on change over time that were specific to each assessment wave (Bartels et 

al., 2004; Burt et al., 2007; Eley et al., 2003). Because nonshared environmental influences tend 

to become more stable with age (Burt et al., 2015; Hopwood et al., 2011), it is possible that our 

study was better positioned to detect stability in the nonshared environment compared to those 

conducted in samples of children and young adolescents (e.g., Bartels et al., 2004; Eley et al., 

2003). The possibility of nonshared environmental influences contributing increasingly to 

stability with age is also consistent with developmental theories of canalization, which posit that, 

as youth begin to shape their own environments, their range of potential outcomes typically 

narrows. In other words, individuals increasingly follow idiosyncratic trajectories in accordance 

with both genetic predispositions and environmental exposures that they themselves may seek 

out (e.g., Turkheimer & Gottesman, 1991). For example, a twin who is parented more harshly 

during preschool may experience difficulty regulating their emotions throughout childhood and 

adolescence and increasingly choose to spend time with peers who have similar difficulties, 

further differentiating the child from their co-twin. That said, this interpretation is not consistent 

with the findings of Burt et al. (2007), which also identified transient effects of the nonshared 

environment between late adolescence and early adulthood. However, Burt and colleagues 

examined diagnostic symptom counts of Antisocial Personality Disorder, a more extreme 

phenotype than the more dimensional ASB assessment examined here (Lahey et al., 2005).  

23 

 
What might be the specific non-shared environmental experiences that underlie stability 

in ASB? One possible non-shared environmental influence is delinquent peer affiliation, which 

has been found to predict growth in ASB throughout adolescence (Eamon, 2002; Gardner et al., 

2008). That said, prior twin work has suggested that twin differences in delinquent peer 

affiliation appear to be a consequence, rather than a cause, of differences in their ASB (Burt, et 

al., 2009). Another possibility centers on aspects of the family environment that, while 

objectively shared by siblings, impact each child in idiosyncratic ways (e.g., siblings respond 

differently to parental divorce) (Goldsmith, 1993). Such familial influences could have an 

enduring impact on ASB development throughout childhood and adolescence. A final possibility 

is differential parenting, which may represent a relatively continuous influence that stably 

differentiates children in the same family. Such considerations are consistent with theoretical 

work positing that “proximal processes”, or reciprocal interactions between the individual and 

their immediate environment, play a critical role in shaping behavioral development 

(Bronfenbrenner, 1988), and empirical work identifying harsh parenting as a risk factor for child, 

adolescent, and young adult ASB in particular (Beauchaine et al., 2005; Conger et al., 1994; 

Gard et al., 2017). Furthermore, studies of within-family differences in parental harshness 

significantly predicted within-pair differences in monozygotic twins’ ASB, both cross-sectionally 

(Burt et al., 2021) and over time (Burt, et al., 2006). Such findings point to parenting as a 

particularly promising target for subsequent studies of the environmental etiology of ASB. 

Notably, however, only environmental exposures unique to each child in a given family 

appeared to impact change in ASB across development, as shared or common family-level 

environmental influences were negligible. While the shared environment has previously been 

found to contribute to continuity in ASB during childhood and early adolescence (Bartels et al., 

2004; Eley et al., 2003), it has not been found to impact ASB development during emerging 

adulthood (Burt et al., 2007). While our inclusion of emerging adulthood could conceivably 

contribute to our null findings for shared environmental influences, we also note that ASB was 

24 

 
assessed using a 5-item screening measure of youth behavior problems. Brief measures often 

have lower reliabilities than do longer measures, an especially salient point here since 

increased measurement error would increase estimates of nonshared environmental effects 

(see Burt (2009) for a discussion of factors affecting detection of shared environmental effects). 

Consistent with the latter, supplemental analyses using participants’ scores on the full CBCL 

Conduct Problems scale (17 items) across the TBED-C and MTwiNS administrations indicated 

that there were significant shared environmental contributions to baseline ASB during middle 

childhood (C variance estimate = .20, p < .05), although shared environmental influences on 

rate of change remained non-significant (C variance estimate = .11). Nevertheless, there were 

also significant, and partially overlapping, genetic and nonshared environmental contributions to 

both intercept and slope for CBCL scores, consistent with our results for the 5-item measure 

(see Table S1.1 and Figure S1.1 in Supplemental Material).  

There are several other limitations to keep in mind when interpreting the results of the 

present study. First, our analyses are not able to clarify the exact duration of nonshared 

environmental contributions to ASB. While the significant overlap in these contributions at 

baseline and over time indicates some degree of continuity, it is unclear whether the influences 

that do not overlap represent transient effects lasting minutes or days, or more enduring effects 

that contribute to systematic change. Second, there was a drop in sample size at ages 11-12 

and 20-22. As our intercept and slope estimates were based on growth curves, however, there 

is relatively little impact of ASB estimates at one particular age on participants’ overall 

trajectories. 

 Next, the SDQ does not delineate aggressive and non-aggressive rule-breaking sub-

types of ASB. This is potentially problematic since these two dimensions of ASB have been 

shown to exhibit distinct etiologies and developmental trajectories (Burt, 2012). Indeed, our 

finding that ASB decreased linearly across development likely indicates that our measure was 

unable to capture the spike in rule-breaking typically seen in studies spanning adolescence 

25 

 
(e.g., Bongers et al., 2004; Windle, 2000). There is thus a need for subsequent research on the 

development of rule-breaking and aggression as separate phenotypes from childhood into 

emerging adulthood, particularly using developmentally sensitive measures that capture 

differences in symptom presentation by age (i.e., heterotypic continuity). However, the items 

included on the SDQ, and in our abbreviated scale from the CBCL, screen for behaviors that are 

typically present, to some degree, throughout early development (e.g., lying, disobedience). 

Moreover, scores on the 17-item Conduct Problems scale on the CBCL also declined linearly 

across the three TBED-C/MTwiNS assessments, which spanned ages 6 to 19, indicating that 

the brevity of our measure likely did not prevent it from capturing age-related trends in ASB 

development in our sample. 

In addition, child sex was entered as a covariate in models including male and female 

participants. Some longitudinal twin studies (e.g., Burt et al., 2007; Eley et al., 2003) have found 

models allowing for sex differences in the etiology of ASB development to fit better than models 

constraining parameters to be equal across sex. However, this pattern of results is generally the 

exception rather than the rule (Burt et al., 2019; Jacobson et al., 2002). Moreover, we note that 

Burt et al. (2007) found few differences between male and female participants in standardized 

parameter estimates. As such, while males evidenced higher levels of ASB at baseline and 

somewhat more decline over time relative to females in our study, we do not expect our overall 

conclusions to differ in models allowing for sex differences in parameter estimates. 

No differences were observed between white participants and those identifying with 

marginalized races/ethnicities in either baseline ASB or change over time in our sample. Given 

the demographics of the State of Michigan, however, there were not sufficient numbers of those 

who identified with any specific marginalized race or ethnicity to model these groups separately. 

There was a significant effect of neighborhood disadvantage, with youth from impoverished 

neighborhoods exhibiting higher levels of ASB at baseline and more rapid decline over time. 

That less privileged youth had higher initial levels of ASB is consistent with a large body of 

26 

 
research demonstrating that familial and neighborhood disadvantage increases risk for nearly all 

youth psychiatric disorders (e.g., Kupersmidt et al., 1995; Leventhal & Brooks-Gunn, 2000). Our 

finding that these youth also desisted more quickly suggests that discrepancies in behavioral 

outcomes by socioeconomic status may decrease with age. Regardless, there is a need for 

further research examining disadvantage in the broader context (e.g., neighborhoods, schools), 

and inequitable structural characteristics (e.g., differences in policing, housing policies) in 

particular, as a predictor of ASB development over time in racially, ethnically, and 

socioeconomically diverse samples. 

Despite these limitations, the present study is the first to examine the genetic and 

environmental etiology of ASB over time in a sample spanning nearly all of childhood, 

adolescence, and emerging adulthood. The key strength of such a study, when incorporating a 

twin design, is its potential to elucidate the genetic and environmental factors contributing to 

human development across multiple stages of the life course. Our study yielded two important 

conclusions. First, genetic factors contributed significantly to both continuity and change in ASB. 

Given the broad age range under study, the genetic contributions to continuity are perhaps 

more noteworthy. Nearly 40% of genetic influences on change throughout development were 

already present at the baseline assessment, which was conducted as early as age 3 in some 

participants. Such findings underscore the importance of genetic influences in shaping ASB 

trajectories. While the specific genetic factors underlying continuity and change are unknown, 

one possibility is that genetic contributions to improved behavioral and emotional regulation are 

activated as youth progress through adolescence, resulting in fewer problem behaviors over 

time. On the other hand, genetic factors underlying dimensions of temperament that are known 

to be predictive of ASB, including negative emotionality and disinhibition, may be among those 

contributing to continuity in ASB, as temperament is both heritable and moderately stable 

throughout development (Ganiban et al., 2008). Future work should seek to contrast and test 

these two possibilities. 

27 

 
Second, the nonshared environmental influences on ASB reflected not only transient 

person-specific environmental influences, but also more enduring influences that overlapped 

across assessment waves. Put another way, environmental influences unique to each child 

within a given family, rather than shared exposures affecting the entire family, were found to be 

important for both stability and change in ASB across early development. Such findings are 

consistent with research indicating the importance of the nonshared environment to behavioral 

outcomes (Plomin & Daniels, 1987). Subsequent studies should seek to identify specific 

nonshared environmental influences that persist over time prior to adulthood.  

28 

 
 
 
Table 1.1. Descriptive statistics: youth ASB scores by age 

Tables 

Age 

Total N  MTP early 

TBED-C  MTP late  MTwiNS  Online  Mean ASB  

0 

0 

0 

0 

6 

18 

12 

38 

36 

58 

62 

146 

150 

96 

78 

18 

2 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

16 

44 

134 

142 

162 

160 

176 

128 

154 

116 

42 

(SD);  
range 
1.91 (1.63); 
0-8 
1.65 (1.49); 
0-8 
1.74 (1.55); 
0-7 
1.49 (1.62); 
0-10 
1.39 (1.54); 
0-10 
1.37 (1.54); 
0-10 
1.15 (1.42); 
0-9 
1.17 (1.47); 
0-7 
1.01 (1.37); 
0-5 
1.09 (1.43); 
0-6 
.88 (1.31); 
0-6 
1.00 (1.56); 
0-10 
.83 (1.42); 
0-9 
.70 (1.26); 
0-9 
.95 (1.43); 
0-9 
.75 (1.09); 
0-5 
.86 (1.39); 
0-9 
.72 (1.15); 
0-5 
.57 (1.07); 
0-5 

3 

4 

5 

6 

7 

8 

9 

10 

11 

12 

13 

14 

15 

16 

17 

18 

19 

20 

21-
22 

288 

284 

286 

858 

584 

682 

556 

476 

118 

116 

200 

290 

312 

260 

254 

146 

156 

116 

44 

288 

284 

286 

196 

132 

254 

128 

34 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

600 

408 

344 

338 

312 

58 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

0 

62 

38 

66 

78 

92 

8 

14 

4 

2 

0 

4 

0 

0 

0 

0 

2 

29 

 
 
 
 
Table 1.2. Sample sizes and correlations in ASB over time 

Time 2  Time 3  Time 4  Time 5 

.49* 
- 

.31* 
.40* 
- 

Time 1 
Time 2 
Time 3 
Time 4 
Time 5 

r 

.30* 
.36* 
.56* 
- 

Mean age 
(SD) 

Mean ASB  
(SD) 

N 

.20 
.46* 
.55* 
.53* 
- 

6.57 (2.09) 
8.63 (2.60) 
15.29 (3.14) 
16.61 (2.16) 
16.79 (.86) 

1.56 (1.61) 
1.28 (1.49) 
.87 (1.34) 
.87 (1.41) 
.48 (.87) 

1996 
1978 
1354 
551 
56 

Note: Bold font and asterisk indicate that the correlation was significantly different than zero at 

p<.05. Time points indicate assessment waves in chronological order, which varies across 

participants (e.g., Time 1 was MTP 1 for some participants, TBED-C for others). 

30 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 1.3. Key parameter estimates from conditional multilevel growth curve model of ASB 

development 

Parameter 

Estimate 

Fixed effects (means) 

Intercept 
Slope 
ADI → intercept 
Sex → intercept 
Ethnicity → intercept 
ADI → slope 
Sex → slope 
Ethnicity → slope 

.667 
-.018 
.004 
.170 
-.035 
-.0002 
-.008 
.001 

S. E. 

.069 
.007 
.001 
.032 
.054 
(.000) 
.003 
.006 

p-value 

<.001 
.010 
<.001 
<.001 
.518 
.001 
.016 
.900 

Random effects (variances) 

Intercept 
Slope 
Intercept-slope covariance 
Residual variance 
Note: Bold font indicates that the estimate was significantly different than zero at p<.05. 

<.001 
<.001 
<.001 
<.001 

.179 
.001 
-.010 
.152 

.015 
.000 
.001 
.006 

Questionnaire type was included as a random, time-varying covariate.  

31 

 
 
 
 
Table 1.4. Standardized variance estimates from univariate ACE model 

Variance estimates 

a 

c 

e 

Univariate 
Intercept 
Slope 
Note: Bold font and asterisk indicate that the estimate was significantly different than zero at 

.37* 
.52* 

.53* 
.45* 

.10 
.03 

p<.05. 

32 

 
 
 
 
 
 
 
Figures 

Mean ASB by Age

e
r
o
c
S
w
a
R
B
S
A

3

2.5

2

1.5

1

0.5

0

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21

Age

Figure 1.1. Age-related change in ASB. 

33 

 
 
 
 
 
 
 
Figure 1.2. Path diagram of a bivariate twin model. The variance in the intercept and the slope 

is partitioned into additive genetic effects (A1 and A2), shared environmental effects (C1 and 

C2), and nonshared environmental effects (E1 and E2). For ease of presentation, this path 

diagram represents one twin in a pair. Standardized path estimates are squared to represent 

the proportion of variance accounted for. 

34 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
CHAPTER 2: UNDERSTANDING DESISTANCE FROM AGGRESSION: AN EMPIRICAL 

INTEGRATION OF PERSON-CENTERED AND VARIABLE-CENTERED APPROACHES 

Abstract 

When leveraged together, variable-centered and person-centered statistical methods have the 

potential to illuminate the factors predicting mental health recovery. However, because extant 

studies have largely relied on only one of these methods, we do not yet understand why some 

youth demonstrate recovery while others experience chronic symptoms. This omission limits our 

understanding of trajectories of physical aggression (AGG) in particular, which are frequently 

characterized by desistance. The present study examined the development of AGG across 

childhood and adolescence via variable-centered and person-centered modeling, with 

neighborhood and family characteristics considered as moderators. Variable-centered results 

indicated a mean-level decline in AGG with age but were less useful for illuminating predictors 

of that decline. Person-centered analyses, by contrast, identified low parent-child conflict and 

high household income as predictors of desistance. Although variable-centered analyses were 

integral to modeling the average AGG trajectory, person-centered techniques proved more 

useful for understanding predictors of desistance.  

35 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Introduction 

Physical aggression (AGG) encompasses a broad spectrum of behaviors that violate the 

personal rights of others, ranging from relatively minor acts (e.g., hitting, kicking) to more 

serious violent crimes (e.g., stabbing, shooting) that bring offenders into contact with the 

criminal justice system. Trajectories of AGG are characterized by both high rank-order stability 

over time and substantial changes with age. A large body of research, for example, has found 

that the frequency of AGG peaks during the preschool years followed by steady declines across 

childhood and adolescence (e.g., Murray et al., 2020; NICHD Early Child Care Research 

Network & Arsenio, 2004), such that desistance increasingly becomes the norm (e.g., Broidy et 

al., 2003). For those youth who do not desist, however, their trajectories often culminate in a 

series of poor adult outcomes (e.g., poor mental and physical health, financial difficulties, 

incarceration; Odgers et al., 2008). 

There is thus a clear need to illuminate how AGG develops over time and to identify 

predictors that interrupt chronic AGG trajectories. A substantial body of research has sought to 

do the former using either variable-centered or person-centered approaches to longitudinal data 

analysis. Variable-centered studies have generally indicated that mean-level increases are 

normative during the preschool years, but that desistance becomes the norm as children 

progress through school. Longitudinal person-centered studies, in turn, have typically identified 

three or four trajectories, with the modal group following a declining trajectory by middle 

childhood (Carroll, Mikhail, & Burt, 2023). Each statistical approach also provides distinct, yet 

complementary, information about the predictors of AGG. For example, most variable-centered 

studies have found males to exhibit greater AGG than females at baseline but comparable rates 

of change, whereas person-centered studies have indicated that males are typically over-

represented in the high-risk trajectories. Likewise, hostile/conflictual family interactions, familial 

socioeconomic disadvantage, and neighborhood deprivation have each been found to predict 

greater baseline levels of AGG in variable-centered models (e.g., Benson & Buehler, 2012; 

36 

 
Olson et al., 2013; Sacco et al., 2015), as well as membership in “chronic” person-centered 

trajectory groups (e.g., Côté et al., 2006; Spano, Rivera, & Bolland, 2010).  

However, all of the above longitudinal studies relied on only one of the two statistical 

approaches (variable-centered or person-centered), with key downstream consequences for our 

understanding of desistance. Variable-centered approaches use growth curve modeling to 

quantify the mean rate of change over time for the entire sample, which is thought to comprise a 

single population (see Figure 2.1a) (McArdle & Epstein, 1987). Individual patterns of continuity 

and change are captured by the intercept, or level of the outcome variable at a particular point in 

time (often at baseline), and the slope, or rate of change. The means of the intercept and slope 

represent the average pattern of development across the full sample, whereas the variances 

represent individual differences in the initial level of the outcome and rate of change, 

respectively (Preacher, Wichman, MacCallum, & Briggs, 2008). For example, a variable-

centered analysis of AGG might find the mean linear slope to be negative and the intercept and 

slope variances to each be significant. These findings would indicate that the average 

participant exhibited a steady decline in AGG with age, but that there were significant between-

person differences in both initial level of AGG and rate of change. 

Person-centered approaches, by contrast, consider the sample under study to comprise 

distinct subgroups that differ by intercept and/or slope. Many person-centered studies of AGG 

have used a technique developed by Nagin (1999), termed semi-parametric mixture modeling or 

latent class growth analysis (LCGA). In this approach, participants are assigned to latent, or 

unobserved, trajectory groups based on their posterior probability of group membership (i.e., 

their likelihood of belonging to a particular group given their scores on the variable of interest) 

(Figure 2.1b). For example, a participant who consistently reports high levels of AGG would 

have a high probability of assignment to a “chronic” trajectory group, whereas a participant who 

reports frequent engagement only at baseline would likely be assigned to a “desisting” 

trajectory. Although groups derived via LCGA are free to vary from one another by intercept and 

37 

 
slope, in most studies, participants within the same group are assumed to follow the same 

trajectory (i.e., intercept and slope variances are fixed within groups). 

Variable-centered and person-centered approaches are each able to provide only a 

partial picture of the development of AGG (Carroll, Mikhail, et al., 2023). The growth curve 

representing average change in AGG in variable centered studies, for example, may not 

adequately capture the developmental heterogeneity observed in key participants. That is, even 

if the mean slope indicates a modest decline over time, some youth might increase in their 

delinquency over time while others may fully desist. Put differently, there may be more 

information in the variance than in the mean slope itself. The primary strength of person-

centered methods lies in their ability to illuminate the slope variance, identifying specific groups 

of youth whose patterns of engagement make them important targets for intervention efforts. 

However, person-centered approaches are critiqued for overextraction, or identifying latent 

classes that do not actually exist in the sample (Bauer & Curran, 2004). Indeed, some 

researchers argue that, because AGG varies continuously across the population, it has no 

qualitatively distinct underlying classes and is best-represented using variable-centered models 

(e.g., Walters & Ruscio, 2013). In turn, proponents of person-centered methods contend that the 

groups serve as an approximation of participants’ trajectories and need not exist in reality to be 

theoretically and empirically useful (Nagin & Tremblay, 2005). Each technique thus has clear 

advantages and important drawbacks. 

In short, although both person-centered and variable-centered approaches provide 

important information, their respective conclusions are necessarily limited by the particulars of 

their analytic approach. One obvious solution to this dilemma is to conduct both sets of analyses 

in the same dataset with the goal of jointly interpreting their findings. What might we gain from 

doing this? Variable-centered studies illuminate normative patterns of development across 

entire samples and are positioned to identify factors distinguishing between initial levels of the 

trait under study, as well as factors predicting rate of change across the entire sample. 

38 

 
However, they are less helpful for examining change among those youth who demonstrate high 

levels of aggression in childhood and subsequently desist. Person-centered techniques are 

uniquely positioned to identify these youth, whose trajectories are of particular interest to most 

clinical psychologists. Moreover, person-centered analyses can not only distinguish between 

youth exhibiting chronic behavior problems and their counterparts who desist, but also identify 

risk and protective factors that predict membership in the desisting versus persisting classes. 

Such methods may also facilitate precision medicine by allowing interventions to be tailored to 

youth following similar trajectories. 

A recent review (Carroll, Mikhail, et al., 2023) began the process of integrating variable-

centered and person-centered approaches through a conceptual review of the developmental 

literatures on antisocial behavior. Of the 124 studies of AGG included in the review, only four 

(3%) examined variable and person-centered techniques simultaneously (Dong, 2016; 

Ehrenreich, Beron, Brinkley, & Underwood, 2014; Ingoldsby, 2003; NICHD Early Child Care 

Research Network & Arsenio, 2004). The results of these studies yielded little information about 

the predictors of desistance from AGG. Two did not examine predictors of the slope variance 

(Dong, 2016; Ehrenreich et al., 2014), and one was under-powered (Ingoldsby, 2003; N = 170). 

Only one study identified specific predictors of desistance, which included higher maternal 

educational attainment and fewer maternal depressive symptoms, but participants were not 

followed past age nine (NICHD & Arsenio, 2004). Moreover, no study to date has employed 

both person-centered and variable-centered techniques to examine predictors of desistance 

from AGG across childhood and adolescence. Such studies would be crucial for identifying 

targets to leverage in intervention efforts for youth demonstrating elevated levels of AGG, who 

comprise a sizeable proportion (~10% to >50%; e.g., Broidy et al., 2003; Nagin, Barker, 

Lacourse, & Tremblay, 2009) of childhood samples.  

Present Study 

The present study sought to illuminate the development of AGG from middle childhood 

39 

 
through late adolescence in a large, population-based sample enriched for neighborhood 

disadvantage. The development of AGG was modeled via a series of variable-centered and 

person-centered analyses, and biological sex, household income, parenting, and neighborhood 

disadvantage were examined as moderators of AGG growth curves and trajectories. We 

examined socioeconomic disadvantage at both the family and neighborhood levels, in light of 

prior research indicating that each form of disadvantage has unique implications for youth 

behavioral outcomes, particularly externalizing (Carroll et al., 2023). Consistent with prior work 

(e.g., Brame, Nagin, & Tremblay, 2001; Lee, Liu, & Watson, 2016), we hypothesized that 1) the 

mean level of AGG would decline across the study period, 2) there would be either three or four 

AGG trajectory groups, and 3) female sex, higher household income, greater parental 

nurturance, lower parent-child conflict, and greater neighborhood advantage would predict lower 

levels of AGG at baseline as well as desistance from AGG across adolescence.  

Transparency and Openness 

The present study was not preregistered. Because of the language in the informed 

consent document at intake, we cannot post the data publicly, but they can be obtained from the 

primary author upon reasonable request. We report how we determined our sample size, all 

data exclusions, all manipulations, and all measures in the study. All procedures were approved 

by the primary author’s institutional review board and are in compliance with the ethical 

standards of the relevant national and institutional committees on human experimentation and 

with the Helsinki Declaration. Child participants provided informed assent, and parents provided 

informed consent for themselves and their children.  

Methods 

Participants 

The population-based Michigan State University Twin Registry (MSUTR) includes 

several independent twin projects. Participants in the current study were drawn from the Twin 

Study of Behavioral and Emotional Development in Children (TBED-C), a study within the 

40 

 
MSUTR. The TBED-C includes a population-based arm (N=528 families), and an ‘under-

resourced’ arm for which inclusion criteria also specified that participating twin families lived in 

neighborhoods with neighborhood poverty levels at or above the Census mean at study onset 

(10.5%) (N=502 families). To recruit families at intake, the Department of Vital Records in the 

Michigan Department of Health and Human Services (MDHHS) identified twins in our age-range 

via the Michigan Twins Project (MTP), a population-based registry of more than 32,000 twins in 

Lower Michigan recruited via birth records. The Michigan Bureau of Integration, Information, and 

Planning Services database was used to locate family addresses no more than 120 miles from 

East Lansing, Michigan through parent drivers’ license information. Pre-made recruitment 

packets were then mailed to parents by the MDHHS. Parents who did not respond to the first 

mailing were sent additional mailings roughly one month apart until either a reply was received 

or up to four letters had been mailed.   

This recruitment strategy for the TBED-C yielded an overall response rate of 57% for the 

at-risk sample and 63% for the population-based sample. Other recruitment and sampling 

details can be found in prior publications (i.e., Burt & Klump, 2019). The two arms of the study 

were analyzed jointly for the current analyses and were 48.7% female and 51.3% male. Across 

the full sample, participants endorsed racial/ethnic identities in the following proportions: white 

(non-Latinx), 82%; Black, 10%; Latinx, 1%; Asian, 1%; Indigenous, 1%; multiracial, 6%. 

However, families in the under-resourced arm, but not the population-based arm, were more 

racially diverse than the local population (e.g., 14% Black and 77% white in the under-resourced 

arm versus 5% Black and 87% white in the population-based arm; 5% Black and 85% white in 

the local area Census).  

Because 90+% of TBED-C families were recruited out of the MTP, we were able to use 

the MTP data to compare families who chose to participate in TBED-C with those who were 

recruited but did not participate. TBED-C families were generally representative of recruited but 

non-participating families. As compared to non-participating twins, participating twins reported 

41 

 
similar levels of conduct problems, emotional symptoms, and hyperactivity (d ranged from -.08 

to .01 in the population-based arm and .01 to .09 in the under-resourced arm; all ns). 

Participating families also did not differ from non-participating families in paternal felony 

convictions (d = -.01 and .13 for the population-based and under-resourced arms, respectively), 

rate of single parent homes (d = .10 and -.01 for the population-based and under-resourced 

arms, respectively), paternal years of education (both d ≤ .12), or maternal and paternal alcohol 

problems (d ranged from .03 to .05 across the two arms). However, participating mothers in 

both samples reported slightly more years of education (d = .17 and .26, both p < .05) than non-

participating mothers. Maternal felony convictions differed across participating and non-

participating families in the population-based arm (d = -.20; p < .05) but not in the under-

resourced arm (d = .02). In short, our recruitment procedures appear to have yielded a sample 

that was representative of both recruited families and the general population of the State of 

Michigan. 

Behavioral data regarding the children’s AGG were collected at up to three time points. 

All 1,030 families were assessed once in middle childhood (ages 6-11; Mage (SD)=8.02 (1.49)) 

as part of the TBED-C (Wave 1). TBED-C participants residing in modestly-to-severely 

disadvantaged neighborhoods (N = 768 families) are currently being recruited for reassessment 

as adolescents up to two times, 18-months apart, through the Michigan Twin Neurogenetics 

Study (MTwiNS). For the present study, the sample sizes at Waves 2 and 3 were 760 and 414, 

respectively.  

Procedure 

TBED-C families completed their intake assessments between 2008 and 2015. 

Assessment teams consisted of two research assistants and at least one paid staff member and 

took 4-5 hours to complete (lunch was provided). Families completed questionnaires, 

interviews, and videotaped interactions during the in-person assessment. The primary caregiver 

(nearly always the mother) participated alongside the twins. Assessments typically took place in 

42 

 
our university laboratory (903 families). If families were unable or unwilling to travel, however, 

assessments took place in participants’ homes (13%). Families with older twins were less likely 

to complete home visits than on-site assessment visits (Cohen’s d=-.24, p=.01), as were 

families that identified as White (d=-.21; p<.05). Similarly, families with fewer financial resources 

were more likely to participate in home visits versus on-site assessments (d=-.32; p=.004). 

However, families completing on-site assessments versus home visits did not differ in maternal 

education, zygosity of the twins, or sex of the twins (ds ranged from -.02 to .16, all ns). The 

MTwiNS follow-up assessments were also completed in-person. The twins and their primary 

caregiver completed an in-person assessment lasting 4-8 hours at either the East Lansing or 

Ann Arbor-based laboratories.  

Measures 

AGG 

Youth AGG was defined via parent report. The primary caregiver (nearly always the 

twins’ mother) completed the 18-item AGG scale from the Child Behavior Checklist (CBCL; 

Achenbach & Rescorla, 2001) at all three waves, rating the extent to which a series of 

statements described the child’s behavior over the past six months using a three-point scale 

(0=never to 2=often/mostly true). Caregiver-informant reports were available for 99% of 

participating twins at Wave 1, 99% at Wave 2, and 90% at Wave 3. One issue with the AGG 

scale from the Achenbach family of instruments is that the items assess not only physically 

aggressive behaviors (e.g., fights, destroys belongings) but also verbal AGG (e.g., argues) and 

emotional dysregulation (e.g., sudden changes in mood). Because the aim of our study was to 

model the development of physical AGG, we constructed a shortened scale of items from the 

Achenbach instruments via a series of item response theory (IRT) analyses (see Supplementary 

Methods in Appendix B). 

Familial Context 

Familial socioeconomic deprivation was assessed via maternal reports of annual 

43 

 
household income at the TBED-C assessment. Household income was measured on a 10-point 

Likert scale (1 = <$10,000 to 10 = >$50,000). 

At the TBED-C assessment, participating mothers and children also completed the 

Parental Environment Questionnaire (PEQ; Elkins et al., 1997), which assesses several 

dimensions of the parent-child relationship using a 4-point Likert scale ranging from “definitely 

false” to “definitely true”. We focused on the Conflict and Nurturance scales (12 items each; α ≥ 

.68 for all scales and informants). The Conflict scale assesses harsh or conflictive parenting 

practices (e.g., “I often criticize my child”), while the Nurturance or Involvement scale assesses 

parental communication, support, and involvement with their child (e.g., “I praise my child when 

he/she does something well”). Items are nearly identical across parent and child versions of the 

PEQ, with minor alterations in wording. Responses were coded so that higher scores indicated 

higher levels of each construct. Scores were averaged across informants to create composite 

reports of conflict and nurturance, respectively. Prior work has found the Conflict and 

Nurturance scales from the PEQ to demonstrate strong internal consistency reliability for 

parents and children and to be highly correlated with scores on the Family Environment Scale, 

which also assesses quality of the parent-child relationship (Elkins et al., 1997).  

Neighborhood Context 

Neighborhood disadvantage was defined via the Area Deprivation Index (ADI; Kind & 

Buckingham, 2018), which comprises 17 indices of Census block group disadvantage (e.g., 

poverty rate, income disparity). Data were weighted according to the factor loadings identified 

by Kind & Buckingham (2018), and weighted variables were summed to create a deprivation 

index score for each Census block group. Families were assigned a percentile indicating the 

level of deprivation in their block group relative to that of all U.S. block groups. Mean ADI scores 

were 57.24 (SD=22.67; range: 2-99), 61.57 (SD=20.94; range: 6-99), and 60.71 (SD=21.62; 

range: 12-99) at the TBED-C and MTwiNS assessments, respectively. These scores 

correspond to a moderate level of neighborhood disadvantage, albeit with considerable variation 

44 

 
across participating families. At the TBED-C assessment, for instance, Census tract poverty 

rates ranged from 0 to 93%, with a mean of 19%.  

Race/Ethnicity 

Given inequalities in exposure to socioeconomic disadvantage across racial/ethnic 

identities (Peterson & Krivo, 2009), race was included as a covariate in all longitudinal analyses, 

coded dichotomously as white/racially marginalized due to the composition of our sample.  

Data Analytic Strategy 

Sum scores for parent-reported AGG were computed based on the items retained in the 

final IRT model (see Supplementary Methods in Appendix B for additional details). These 

scores were log-transformed to adjust for positive skew and were subsequently examined as 

the outcome measures in the variable-centered and person-centered analyses. Continuous 

covariates (i.e., income, parenting, and neighborhood disadvantage) were standardized for ease 

of interpretation. 

Variable-Centered Analyses 

Analyses were conducted in Mplus 8.4 (Muthén & Muthén, 1998-2019) using a robust 

maximum likelihood estimator and full information maximum likelihood estimation (FIML) to 

account for missing data on the outcome. We examined phenotypic changes in the frequency of 

AGG over time across the full sample via latent growth curve modeling, which identifies within-

person changes in the outcome of interest as well as between-person variation in these 

changes (McArdle & Epstein, 1987). First, unconditional models were run to determine whether 

engagement in AGG increased, decreased, or remained constant with age. Age was examined 

in lieu of assessment wave as the metric of time given the wide range of ages included at each 

assessment (see Table S2.1). The definition variable approach was used to scale time in years, 

with the factor loadings constrained to vary by age (Sterba, 2014). Because the definition 

variables must be complete for the model to be estimated correctly, placeholder values were 

imputed to represent age for participants with missing data (Grimm, Ram, & Estabrook, 2017). 

45 

 
Specifically, the mean number of years that elapsed between waves was added to participants’ 

ages at Wave 1 if data were missing at Waves 2 and/or 3. Age was available for all participants 

at Wave 1. Imputed values had no impact on estimation, as missing data on the definition 

variables was connected with missing data on the outcomes.  

Intercepts were centered to represent AGG at age six, the youngest age at the first 

TBED-C assessment wave. The slope factor represented annual change. Because participants 

were assessed on only three occasions, we did not examine non-linear growth. Initial analyses 

compared the respective fits of an intercept-only (“no-growth”) model and a model allowing for 

linear growth. The former had three parameters (intercept mean, intercept variance, and 

residual variance), and the latter had six (intercept and slope means, intercept and slope 

variances and their covariance, and residual variance). Model fit was evaluated with four 

likelihood-based indices: the chi-square difference test (Bollen, 1989), Akaike information 

criterion (AIC; Akaike, 1987), Bayesian information criterion (BIC; Raftery, 1995), and sample-

size adjusted Bayesian information criterion (SABIC; Sclove, 1987). For all indices, lower values 

indicated better model fit. In the unconditional models, the CLUSTER command was used to 

account for the nesting of twins within families. 

We selected the best-fitting model for the conditional growth curve analyses. AGG was 

fit conditional upon sex, race, income, conflict, nurturance, and ADI scores, all of which were 

examined as time-invariant predictors. Because Mplus employs listwise deletion for missing 

covariates, all available data were analyzed in the conditional growth curve model by using 

multiple imputation (Little & Rubin, 2019), with ten imputed datasets. As a result, estimates of 

the parameters and standard errors reflect the uncertainty that is due to missing data. In Mplus, 

it is not possible to include the CLUSTER command while using multiple imputation. Thus, in 

the conditional LGM, the data were structured to be in “family-wide” format, with parameters 

constrained to equality across both twins in a pair. The conditional LGM also accounted for the 

difference in within-pair resemblance for monozygotic (identical) and dizygotic (fraternal) twins, 

46 

 
as identical twins are expected to resemble one another more so than fraternal twins on virtually 

every trait (Turkheimer, 2000).  

Person-Centered Analyses 

Latent class growth analyses were conducted in Mplus 8.4 using a robust maximum 

likelihood estimator and FIML to account for missing outcome data. We examined participants’ 

growth in AGG as a function of their age at each assessment wave. We began with a one-class 

model and progressed through to a six-class model, consistent with the steps outlined by 

Muthén (2004) and with the approach employed in prior studies examining trajectories of 

antisocial behavior (e.g., Odgers et al., 2008). All models allowed the groups to differ from one 

another by intercept and slope but constrained within-group intercept and slope variances to 

zero. The optimal model was determined via consideration of two fit indices, the AIC and BIC, 

as well as the interpretability of the class solution (i.e., the percentage of the sample assigned to 

each class). Biological sex, race, income, parenting, and ADI were subsequently examined as 

time-invariant predictors of class membership, using ten imputed datasets to account for 

missing values. Notably, it is not possible to account for clustering when examining age as the 

metric of time in Mplus mixture models, nor was it feasible to examine the data in “family-wide” 

format because both twins in a pair would be assigned to the same trajectory by default. Thus, 

we did not account for clustering in our main analyses but ran supplemental analyses with one 

twin randomly selected per family, discussed in greater detail later in the manuscript. 

Descriptives 

Results 

Descriptive statistics and correlations for the AGG sum scores are shown in Table 2.1. 

Participants evidenced moderate rank-order stability in AGG over time. Moreover, paired-

sample t tests indicated that AGG decreased in frequency from Wave 1 to Wave 2 (t(747) = 

13.78, p < .001), and from Wave 2 to Wave 3 (t(366) = 2.13, p < .05). At the first wave, males 

had higher levels of AGG than females (Cohen’s d = .285; p < .001); although in the expected 

47 

 
 
 
direction, sex differences were not significant at Waves 2 or 3. In addition, ADI, income, and 

nurturance were significantly correlated with AGG at the first two time points, but not at the third. 

Conflict was significantly correlated with AGG at all waves. All correlations were in the expected 

direction. (See Table S2.1 for descriptives by age, rather than wave). 

Variable-Centered 

Unconditional models 

Initial growth curve analyses tested the hypothesis that the frequency of AGG decreased 

with age by comparing the respective fits of an intercept-only model and a linear growth model. 

The linear growth model fit better than the intercept-only model according to all fit indices (see 

Table 2.2) and was thus selected as the final unconditional model. At age six, the average 

participant had a score of .810 on AGG (SE = .022, p< .001). On average, scores decreased by 

.051 points per year (SE = .003, p< .001), consistent with hypotheses. There was significant 

between-person variation in baseline level (b = .379, SE = .030, p< .001) and in rate of change 

(b = .003, SE = .001, p< .001). The estimated covariance between the intercept and slope was -

.026 (SE = .004, p< .001), indicating that participants with higher AGG scores at baseline 

declined more rapidly over time.  

Conditional model 

Sex, race, income, conflict, nurturance, and ADI were subsequently examined as 

predictors of continuity and change in AGG. As hypothesized, male sex, low income, parent-

child conflict, and ADI predicted greater AGG at baseline (see Table 2.2). Race and nurturance 

predicted neither baseline AGG nor rate of change. The only significant predictors of the slope 

were sex and parent-child conflict, with males and those with greater conflict demonstrating a 

more rapid age-related decline. 

Person-Centered 

Unconditional models 

Initial analyses compared the respective fits of a one-class model through to a six-class 

48 

 
model to test the hypothesis that there would be either three or four distinct AGG trajectories. 

Model fit improved as the number of classes increased, as shown in Table 2.3. However, the 

four through six-class solutions each contained groups that did not appear to follow 

meaningfully distinct trajectories. For example, the four-class solution comprised two stable 

trajectories, one of which had a somewhat higher baseline level than the other, whereas the five 

and six-class solutions each contained two groups that followed desisting trajectories. The 

three-class solution was thus selected as the optimal model based on fit and clinical relevance. 

As shown in Figure 2.2, the three-class model comprised a low, stable trajectory representing 

the majority (53%) of the sample, as well as two trajectories characterized by high levels of 

AGG at baseline. One of the latter trajectories (33.2% of the sample) followed a desisting 

pattern with age, whereas the other group (14.1%) continued to aggress at high levels.  

Conditional model 

Male sex, high ADI, low income, and high conflict predicted membership in the persisting 

group relative to the low, stable group. Likewise, income and conflict distinguished between 

desisting and persisting youth. Namely, youth with high levels of AGG in childhood who resided 

in more affluent households, as well as those exposed to less conflict, were more likely to desist 

from AGG over the course of adolescence (see Table 2.3 for conditional model results and 

Table 2.4 for trajectory group characteristics). 

Interpretive Integration 

Variable-centered and person-centered analyses both indicated a decline in AGG with 

age. Moreover, male sex, parent-child conflict, low household income, and neighborhood 

disadvantage all predicted elevated AGG at baseline, evidenced by their prediction of the 

intercept (variable-centered) and their distinction between the normative group and the 

elevated, persistent group (person-centered). However, variable-centered and person-centered 

results differed with respect to predictors of the rate of change. In the former, only sex and 

mother-reported conflict predicted the slope, and results were not in the expected direction (i.e., 

49 

 
male sex and greater conflict predicted faster decline with age). Such findings may reflect a 

statistical artifact, as males and those with greater conflict evidenced higher baseline levels of 

AGG and thus had more room to decline over time. In the person-centered analyses, by 

contrast, lower levels of conflict, as well as familial affluence, predicted desistance from AGG 

among youth demonstrating elevated engagement at baseline. As the desisting and persisting 

trajectory groups differed primarily by slope, such findings implicate parent-child conflict and 

familial disadvantage as predictors of the slope variance. Moreover, they point to limitations of 

variable-centered approaches in understanding heterogeneity in rates of change.  

Discussion 

The aims of our study were to elucidate the predictors of desistance from AGG across 

middle childhood and adolescence by integrating findings across two distinct statistical modeling 

approaches. Consistent with our hypotheses, the modal pattern was one of declining AGG with 

age. That is, variable-centered analyses revealed a mean-level decline in AGG across the study 

period, and all three trajectories identified via person-centered modeling followed either stable 

or decreasing patterns. These findings are consistent with a large body of literature indicating 

that engagement in AGG typically begins to decline by school entry and continues to do so 

throughout adolescence (e.g., Broidy et al., 2003; Ehrenreich et al., 2014), facilitated by age-

related improvements in emotion regulation and behavioral inhibition (Burt, 2012). The number 

and shape of the trajectory groups we identified were also largely consistent with findings from 

studies spanning middle childhood and adolescence, with a modal trajectory characterized by 

little AGG and two higher-risk trajectories with differing degrees of desistance (e.g., Becht et al., 

2016; Maughan et al., 2000).  

Although findings were largely consistent across variable and person-centered analyses, 

each approach provided unique information about the development of AGG, particularly in 

relation to associations with the predictors. The variable-centered analyses revealed that 

baseline levels of AGG differed as a function of sex, mother-child conflict, familial 

50 

 
socioeconomic status, and neighborhood deprivation. However, only sex and mother-child 

conflict were observed to predict the slope, and with small effect sizes. Based on the variable-

centered findings, one might conclude that youth largely followed similar patterns of declining 

AGG with age, albeit with some variation between-persons in rate of decline. The person-

centered analyses, however, painted a different picture. The trajectories differed not only by 

baseline engagement but also in their developmental patterns, with one group accounting for 

most of the age-related decline observed across the full sample.  

Perhaps most noteworthy among the person-centered findings is that a sizeable 

proportion of participants with elevated AGG at baseline had desisted by early adulthood. This 

finding raises a question that person-centered approaches are uniquely positioned to answer, 

namely, which specific factors interrupt trajectories of elevated AGG? In contrast to the variable-

centered findings, the factors predicting patterns of change in the mixture models did so in a 

way that was consistent with theory. Specifically, familial socioeconomic advantage, as well as 

parent-child relationships characterized by little conflict, predicted desistance among youth 

demonstrating elevated AGG at baseline. Not only do these findings suggest potential targets 

for intervention, but they also point to the utility of person-centered approaches to the study of 

AGG (and other forms of psychopathology) even within dimensional data. Because variable-

centered analyses examine development across entire samples, they are not positioned to 

extract subgroups that differ widely in their patterns of change, or to identify specific 

environmental factors predicting such different patterns. Similarly, although person-centered 

models of antisocial behavior subdivide a continuous trait into distinct categories (e.g., Walters 

& Ruscio, 2013), the present findings indicate that such approaches can yield information about 

developmental patterns in continuously distributed data that variable-centered techniques 

cannot. 

We next discuss each of the salient predictors in more detail. As hypothesized, 

socioeconomic advantage predicted less engagement in AGG. In the growth curve model, youth 

51 

 
from wealthier households and non-deprived Census block groups demonstrated lower levels of 

AGG at baseline than their peers from impoverished backgrounds. Likewise, in the person-

centered analyses, high income and low ADI both predicted membership in the stable/low AGG 

trajectory relative to the persistent one. These findings are consistent with theoretical work 

pointing to numerous environmental contexts (e.g., family, school, neighborhood) that impact 

child development (Bronfenbrenner, 1988). They also extend prior, cross-sectional research 

implicating socioeconomic disadvantage as a multi-faceted construct, with non-interchangeable 

effects from the family environment and the broader neighborhood context (Carroll et al., 2023; 

Kupersmidt et al., 1995). Youth who reside in homes and neighborhoods equipped with 

adequate resources are thus more likely to exhibit low levels of AGG. Moreover, even if they do 

develop AGG during childhood, those from affluent homes are far more likely to desist than their 

peers from impoverished backgrounds.  

Aspects of the parent-child relationship also predicted trajectories of AGG. Specifically, 

low levels of mother-child conflict predicted low AGG at baseline in the variable-centered 

analyses, consistent with prior studies of parent-child conflict and youth behavior problems (El-

Sheikh & Elmore-Staton, 2004; Xu et al., 2021). Likewise, in the person-centered analyses, low 

levels of conflict predicted both low and desisting AGG trajectories relative to persistent AGG. 

Notably, conflict was one of only two variables that predicted the slope of the latent growth 

curve, and results were inconsistent with those of the person-centered analyses, such that the 

former found higher levels of conflict to predict a faster decline and the latter found lower levels 

of conflict to predict desistance. The unexpected findings in the variable-centered analyses may 

reflect regression to the mean among participants demonstrating high levels of AGG at 

baseline. That is, participants reporting higher levels of conflict also demonstrated greater 

baseline AGG and thus had more room to decline with age. The prospect of certain variable-

centered findings reflecting regression to the mean artifacts is consistent with prior simulation 

results. Marsh & Hau (2002), for example, simulated test score data in which schools differed by 

52 

 
average achievement at baseline but followed identical growth trajectories. Nevertheless, 

estimates from multilevel growth curve analyses indicated that schools with lower initial 

averages demonstrated greater growth in scores over time. An analogous pattern was observed 

in the present data, namely that participants with higher initial AGG scores, including those with 

elevated parent-child conflict, appeared to decline more with age.  

By contrast, when we examined predictors of trajectory group membership, results for 

conflict were in the expected direction, with lower conflict predicting both consistently low and 

desisting trajectories. Because participants in the desisting group had greater conflict than those 

in the normative group, who had little room to decline with age, higher conflict was associated 

with more rapidly declining AGG in analyses that collapsed across the full sample. Although 

reflective of the high levels of conflict among participants with elevated AGG at baseline, 

variable-centered findings thus provided a somewhat misleading picture of the factors predicting 

the slope variance. Integration of the variable-centered and person-centered results 

demonstrates that greater conflict predicts elevated AGG at baseline and continued 

engagement in AGG across adolescence.  

Contrary to our hypotheses, however, nurturance did not serve as a protective factor. 

Prior work has found parental nurturance to have a small association with physical AGG, but at 

different ages for boys and girls (Arim, Dahinten, Marshall, & Shapka, 2009) or only in 

conjunction with early pubertal timing in girls (Mrug et al., 2008). Overall, the present findings 

suggest that negative aspects of the parent-child relationship may be more closely tied to the 

development of AGG than are positive aspects of the parent-child relationship.  

As hypothesized, males demonstrated higher levels of AGG at baseline and were more 

often assigned to the persistent trajectory group relative to the normative group. Sex also 

predicted the slope of the growth curve, with males exhibiting a somewhat faster decline in AGG 

with age, which may again reflect regression to the mean (although males did not exhibit lower 

levels than females on average prior to age 18; see Figure S2.1). Notably, however, sex did not 

53 

 
distinguish between desisting and persisting trajectory group membership, as both groups were 

predominantly male. Such findings are fully consistent with the broader literature, which has 

found AGG to be highly sex-specific throughout early development, with some studies 

identifying up to 15-fold higher rates of AGG for males relative to females (e.g., Berkout, Young, 

& Gross, 2011). 

Limitations 

Our results should be interpreted in light of several limitations. First, participants were 

predominantly white. Although our sample is socioeconomically diverse and representative of 

the racial/ethnic makeup of the State of Michigan, future studies should examine whether our 

findings generalize to youth from more diverse racial/ethnic backgrounds as well. Next, 

participants were assessed on up to three occasions, meaning it was not possible to model non-

linear changes in AGG. This may be particularly salient when modeling the development of 

AGG for adolescents in under-resourced contexts, whose engagement often fluctuates more 

than that of their wealthier peers (e.g., Ehrenreich et al., 2014; Karriker-Jaffe, Foshee, Ennett, & 

Suchindran, 2008). Our study was not positioned to detect such fluctuations, although our 

finding that neighborhood and familial deprivation both predicted persistent AGG is consistent 

with prior findings that youth exposed to disadvantage are less likely to follow a steadily 

declining AGG trajectory. 

Next, our measure assessed the frequency, not the severity, of AGG, and thus did not 

detect the age-related increases in severity reported in prior research (e.g., Copeland, Miller-

Johnson, Keeler, Angold, & Costello, 2007). In addition, items assessed behaviors of relatively 

low severity (e.g., hot temper, mean) rather than violent crime, which may explain why we did 

not identify a “late-onset” or “adolescent-increasing” group in the person-centered analyses. 

Studies identifying groups with onset after childhood typically examined serious acts of violence 

(e.g., shooting, stabbing) as the outcome and did not assess for the presence of less serious 

forms of AGG (e.g., Mata & van Dulmen, 2012; Tung & Lee, 2017). In short, our finding that 

54 

 
AGG was most frequent at study onset (i.e., age 6) is consistent with the broader AGG 

literature, which has found that AGG rarely first emerges during adolescence or adulthood 

(Carroll, Mikhail, et al., 2023). Future studies are needed to elucidate associations between 

familial and neighborhood risk factors and the specific types of AGG behaviors comprising very 

high-risk trajectories. 

Next, we focused on maternal reports of AGG, as they were available at all waves. 

Although such an approach eliminates confounding due to differences between participants in 

the number of available informants (i.e., teacher-report data were available for most, but not all, 

participants at the first two waves), there are inherent limitations to focusing on a single 

informant. For example, parents may not be aware of the extent of their children’s AGG, 

particularly during adolescence, given age-related increases in independence and ability to 

conceal problem behaviors (Burt, 2012). In addition, parent reports largely reflect behaviors 

occurring in the home and may not adequately capture AGG in other settings (e.g., school; 

Achenbach, McConaughy, & Howell, 1987). We thus computed mean scores based on data 

from all available informants (i.e., mother, teacher, and self-report) and examined them as the 

outcome in a series of sensitivity analyses. As shown in Tables S2.6 and S2.7 and Figure S2.2, 

results were fully consistent with those from the main analyses. In particular, the three-class 

LCGA solution yielded a “desisting” and a “persisting” trajectory group, with membership in the 

former predicted not only by low conflict and high income, but also by low ADI. In the main 

analyses, ADI was marginally significant as a predictor of desistance relative to persistence 

(p<.10). Additional studies are needed to elucidate the role of neighborhood advantage in 

predicting desistance from AGG. Regardless, our primary results do not appear to be specific to 

parent-reported AGG but instead persist to multi-informant composite reports as well. 

Another methodological limitation concerns the nested structure of the data. As noted 

earlier, we could not include the CLUSTER command in the conditional LGM or in any of the 

person-centered models. We accounted for clustering in the conditional LGM by examining the 

55 

 
data in “family-wide” format. As a verification, we also ran the conditional LGM with the data in 

“long” format, without controlling for nesting. Results were fully consistent with those reported in 

the main analyses (see Table S2.8). To verify results from the mixture models, in which we were 

unable to control for cluster effects, we ran the unconditional and conditional models with one 

twin per family randomly selected (see Table S2.9). Consistent with the results in the full 

sample, the three-class solution comprised a normative group following a consistently low 

trajectory and two other groups differentiated by their degrees of desistance. Results from the 

conditional model also replicated those in the full sample, with income and conflict distinguishing 

between desisting and persisting classes. We thus do not believe that our results were biased 

by the limitations of the software in accounting for clustering.  

A limitation specific to the mixture models was that the average posterior probability was 

relatively low for the persisting trajectory. In mixture modeling, each participant is assigned to 

the group for which their posterior probability of membership is highest. Average posterior 

probabilities are thus a measure of classification accuracy, as higher values indicate greater 

certainty in class assignment. The average probability for the persisting group was .65, 

somewhat below the recommended threshold of .70 (Nagin, 2005). This is likely related to the 

study’s accelerated longitudinal design, with incomplete follow-up at the second and third waves 

of data collection. When the sample was restricted to those with complete data at all waves 

(N=370), average posterior probabilities for all three groups were > .85 (both the class solution 

and the percentage of the sample comprising each group were fully consistent with that 

observed in the full sample). In addition, average posterior probabilities for the desisting and 

low/stable trajectories were .76 and .92, respectively, in the full sample. The three-class solution 

thus assigned participants to trajectories with a reasonable degree of confidence, despite the 

planned missingness at later waves. 

Lastly, all predictors of AGG trajectories were modeled as time-invariant based on 

scores at the middle childhood assessment, when sample size was largest. Although beyond 

56 

 
the scope of the present study, changes in neighborhood conditions, neighborhood residence, 

household income and/or parenting behaviors may elicit changes in AGG. In the case of 

parenting, the association may be bidirectional (i.e., changes in AGG may evoke specific 

parenting behaviors, as suggested in prior work (Klahr, Klump, & Burt, 2014; Narusyte et al., 

2011)). Future studies are needed to examine parallel trajectories of family characteristics and 

youth AGG, particularly to clarify how parenting and AGG develop in tandem.  

Implications 

Despite these limitations, the present study has several important implications. First, our 

findings underscore the importance of leveraging variable-centered and person-centered 

approaches when studying developmental trajectories, even when the trait in question appears 

to be dimensional. Variable-centered approaches are integral to understanding the overall, 

sample-level pattern as well as the predictors of the intercept variance. Given the heterogeneity 

in trajectories of psychopathology within typical samples, however, the magnitude and direction 

of change over time cannot be fully captured by a single slope estimate. Nor can we fully 

understand the predictors of the slope variance when collapsing across the entire sample. Thus, 

after determining via growth curve modeling that there is significant variance between persons 

in rate of change, one good solution would be to turn to person-centered techniques to 

understand the sources of that variance. By identifying discrete trajectories that differ not only 

by baseline level but also by direction/rate of change, person-centered techniques can both 

model the heterogeneity in participants’ slopes and examine the predictors of this heterogeneity 

in a more nuanced way than can be done in variable-centered approaches.  

In the present study, for example, the unconditional variable-centered analyses 

established that there was significant variance in participants’ slopes, but the conditional 

analyses suggested that socioeconomic status did not significantly explain any of this variance, 

a finding that is highly inconsistent with prior research (Carroll, Mikhail, et al., 2023). Moreover, 

low parent-child conflict was implicated as a risk factor for continued AGG, despite predicting 

57 

 
lower levels of AGG at baseline. It was only through the conditional mixture models that we 

were able to identify familial advantage and low parent-child conflict as specific predictors of 

desistance over time, and in a way that was consistent with theory and prior research. In short, 

the integration of the two approaches yielded a far more complete picture of the development of 

AGG, as well as a more logical conceptualization of the factors predicting interindividual 

variation, than either approach provided by itself.  

Second, the present study advances our understanding of youth AGG as a 

developmental phenomenon unfolding within multiple contexts. Individual, familial, and 

neighborhood characteristics all emerged as important predictors of continuity and change in 

AGG. In particular, our findings implicate both the family and the neighborhood as critical 

contexts for the emergence of AGG, given that mother-child conflict, familial poverty, and 

neighborhood deprivation all contributed significantly to AGG trajectories. Moreover, among 

those youth with high levels of AGG at baseline, familial characteristics predicted which youth 

eventually desisted. The distinction between desistance and persistence among those with high 

levels of a given psychopathology is of great clinical importance, given that all who present for 

treatment presumably demonstrate elevated symptoms at baseline. Successful courses of 

treatment would thus be reflected in trajectories of desistance, meaning that factors promoting 

desistance are important to identify and leverage in targeted interventions.  

The present findings specifically point to mother-child conflict as a potential treatment 

target for reducing youth AGG, consistent with the demonstrated efficacy and effectiveness of 

parent management training (PMT). PMT, a structured intervention in which parents are taught 

behavioral strategies to promote their children’s prosocial behavior and discourage conduct 

problems, has proven effective in reducing aggressive, disruptive, and noncompliant behaviors 

in children and adolescents (DeGarmo, Patterson, & Forgatch, 2004; DeGarmo & Forgatch, 

2005; Hagen, Ogden, & Bjørnebekk, 2011; Kjøbli, Hukkelberg, & Ogden, 2013). Moreover, 

studies of treatment mechanisms have found increases in effective parenting (e.g., consistent 

58 

 
discipline, positive involvement) to mediate the effects of PMT (DeGarmo et al., 2004; DeGarmo 

& Forgatch, 2005; Hagen et al., 2011). One goal for future work should be to determine to what 

extent parent-child conflict specifically serves as a mediator for the effects of parent training, as 

well as whether findings are consistent across informants (i.e., mother, father, and child reports 

of conflict). 

Socioeconomic advantage within the home also appeared to interrupt trajectories of 

elevated AGG, consistent with prior work that has found exposure to deprivation to be a potent 

predictor of externalizing behaviors in particular (Carroll et al., 2023; Snyder, Young, & Hankin, 

2019). These findings underscore the implications of socioeconomic disadvantage for youth 

mental health not only at a single timepoint but throughout the early developmental period. In 

more affluent households, for example, healthy development may be facilitated by greater 

access to resources, such as recreational facilities, structured activities, and, notably, mental 

health treatment. Indeed, the markers of desistance identified in the present study may partially 

reflect treatment effects, as youth from wealthier backgrounds who exhibit high levels of AGG 

may be better able to access treatment than their counterparts from disadvantaged homes. 

Future studies should examine access to mental health treatment as a potential mediator of the 

association between familial (dis)advantage and desistance from youth AGG. Studies are also 

needed to determine the extent to which changes in socioeconomic status over time predict 

trajectories of AGG.   

Lastly, future research should employ variable and person-centered techniques to 

illuminate the effects of (dis)advantage on trajectories of other externalizing behaviors, such as 

non-aggressive rule-breaking (RB). In contrast to AGG, RB follows a developmental pattern 

characterized by increasing engagement during adolescence, as well as greater fluctuations 

and lower rank-order stability (Burt, 2012). One might thus expect to identify trajectories of RB 

that differ not by degree of desistance but rather by the extent of their increase across 

adolescence. Leveraging variable and person-centered methods would allow us to determine 

59 

 
the risk factors that predict large increases in RB over time, which would represent promising 

targets for treatment efforts. 

60 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 2.1. Descriptive statistics and correlations 

Tables 

1. ADI 
2. Income 
3. Conflicta 
4. Nurturancea 
5. AGG T1b 
6. AGG T2b 
7. AGG T3b 

Mean (SD) 

Range 
N 

1. 
- 
-.40 
.02 
-.15 
.15 
.08 
.04 

2. 

- 
-.01 
.13 
-.16 
-.11 
-.06 

3. 

4. 

5. 

6. 

7. 

- 
-.37 
.35 
.20 
.17 

- 
-.16 
-.09 
-.01 

- 
.39 
.39 

- 
.62 

- 

57.25 
(22.68) 
2-99 
2010 

8.24 
(2.68) 
1-10 
2010 

20.66 
(4.63) 
12-39 
2041 

41.45 
(3.53) 
22-48 
2041 

1.58 
(1.87) 
0-8 
2040 

.69 
(1.23) 
0-8 
759 

.58 
(1.13) 
0-6 
372 

Note: Bold font indicates p<.05. aMean scores were computed based on maternal and child 

reports. bSum scores were computed based on maternal reports on eight CBCL items. 

61 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 2.2. LGM fit statistics and parameter estimates 

Model 
Linear growth 
Means model 

Unconditional LGM model fit statistics 
-2lnL 
5667.38 
6119.52 

χ2 (df) 
- 
452.14† (3) 

AIC 
5679.38 
6125.52 

BIC 
5713.17 
6142.41 

SABIC 
5694.10 
6132.88 

Conditional LGM parameter estimates 

Parameter 

Estimate 

Mean 
Variance 
ADI  
Income  
Conflict  
Nurturance  
Sex  
Race  

Mean 
Variance 
ADI 
Income  
Conflict  
Nurturance  
Sex  
Race  

.692 
.311 
.055 
-.068 
.250 
.017 
.210 
.016 

-.046 
.003 
-.003 
.002 
-.016 
.001 
-.018 
.005 

S.E. 
Intercept 
.055 
.024 
.023 
.025 
.019 
.017 
.038 
.058 
Slope 
.008 
.001 
.004 
.004 
.003 
.003 
.006 
.008 
Covariances 
.003 

p-value 

<.001 
<.001 
.017 
.006 
<.001 
.325 
<.001 
.784 

<.001 
<.001 
.392 
.551 
<.001 
.849 
.002 
.548 

-.022 

<.001 

.025 
.024 

.274 
.163 

Intercept with slope 
Intercept 1 with 
Intercept 2 
          MZ 
          DZ 
Slope 1 with Slope 2 
          MZ 
          DZ 
Intercept 1 with 
Slope 2 
            MZ 
            DZ 
Residual variance 
Note: Bold font indicates p<.05. †Significant change in chi-square at p<.05.  

<.001 
<.001 
<.001 

-.023 
-.013 
.135 

.003 
.003 
.012 

<.001 
.002 

<.001 
<.001 

.003 
.001 

.001 
.000 

62 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 2.3. LCGA fit statistics and predictors of group membership 

Unconditional LCGA model fit statistics 

AIC 

BIC 

Smallest group (% of 
sample) 

# of groups 
1 
2 
3 
4 
5 
6 

5983.49 
5467.89 
5197.92 
5077.70 
4987.81 
4931.54 

6000.36 
5501.64 
5248.55 
5145.21 
5072.19 
5032.80 

100.0 
38.0 
14.1 
8.2 
5.6 
5.4 

Conditional LCGA parameter estimates from 3-group model 
Persistent v. low/stable 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 
Note: Bold font indicates p<.05. 

Estimate 
.310 
-.396 
1.241 
.147 
.717 
.250 

S.E. 
.102 
.086 
.112 
.100 
.180 
.224 
Persistent v. desisting 
S.E. 
.119 
.095 
.112 
.118 
.218 
.259 
Desisting v. low/stable 
S.E. 
.078 
.079 
.084 
.075 
.142 
.185 

Estimate 
.112 
-.022 
.873 
.019 
.475 
.126 

Estimate 
.198 
-.374 
.369 
.128 
.242 
.123 

p-value 
.002 
<.001 
<.001 
.144 
<.001 
.265 

p-value 
.094 
<.001 
.001 
.277 
.266 
.635 

p-value 
.154 
.780 
<.001 
.805 
.001 
.494 

63 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 2.4. Characteristics of AGG trajectory groups 

N 
% male 

ADI 

Income 

Conflict 

Nurturance 

AGG T1 

AGG T2 

AGG T3 

AGG T1 (log-
transformed) 
AGG T2 (log-
transformed) 
AGG T3 (log-
transformed) 

Class 1: low, stable  Class 2: desisting 

1079 
46.2 

681 
54.3 

Means (SD) 
[Range] 

Class 3: persisting 
290 
64.5 

55.62 (22.22)  
[2-99] 
8.42 (2.61)  
[1-10] 
19.23 (4.12) 
[12-34.5] 
41.79 (3.49) 
[27-48] 
.25 (.44)  
[0-2] 
.24 (.58)  
[0-5] 
.20 (.54)  
[0-3] 
.17 (.30) 
[0-1.10] 
.15 (.33) 
[0-1.79] 
.12 (.31) 
[0-1.39] 

57.15 (23.28)  
[2-99] 
8.41 (2.52) 
[1-10] 
21.82 (4.40) 
[12.5-36.5] 
41.23 (3.51) 
[22-48] 
2.41 (1.24)  
[0-8] 
.45 (.82)  
[0-6] 
.34 (.71)  
[0-4] 
1.17 (.34) 
[0-2.20] 
.26 (.42) 
[0-1.95] 
.21 (.38) 
[0-1.61] 

63.62 (21.70)  
[5-99] 
7.22 (3.06)  
[1-10] 
23.34 (4.96) 
[12.5-39] 
40.68 (3.61) 
[28-48] 
4.56 (1.87)  
[0-8] 
2.68 (1.54)  
[0-8] 
2.48 (1.60)  
[0-6] 
1.65 (.40) 
[0-2.20] 
1.22 (.40) 
[0-2.20] 
1.13 (.51) 
[0-1.95] 

64 

 
 
 
 
 
Figures 

Variable-centered: growth curve model

8

9

10

11

Age

12

13

14

Participant 1

Participant 2

Participant 3

Participant 4

Sample mean

Participant 5

Person-centered: mixture model

8

9

10

11

Age

12

13

14

Group 1: low, stable

Group 2: moderate, desisting

Group 3: high, increasing

G
G
A

4
3.5
3
2.5
2
1.5
1
0.5
0

G
G
A

4
3.5
3
2.5
2
1.5
1
0.5
0

a.

b.

Figure 2.1. Variable-centered (a) and person-centered (b) approaches to modeling the 

development of AGG in a hypothetical childhood sample. The variable-centered approach 

identified a mean-level trend across the entire sample, whereas the person-centered analysis 

extracted three distinct trajectory groups. 

65 

 
 
 
 
 
 
 
 
 
1.8

1.6

1.4

1.2

1

0.8

0.6

0.4

0.2

0

G
G
A

Group 1: low, stable
(52.6%)

Group 2: desisting
(33.2%)

Group 3: persisting
(14.1%)

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21

Age

Figure 2.2. Unconditional three-group model of AGG trajectories (person-centered). 

66 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
CHAPTER 3: UNDERSTANDING HETEROGENEITY IN TRAJECTORIES OF RULE-

BREAKING VIA INTEGRATION OF VARIABLE-CENTERED AND PERSON-CENTERED 

METHODS 

Abstract 

Trajectories of antisocial behavior, including non-aggressive rule-breaking (RB), are 

characterized by considerable heterogeneity between persons. The best-known framework for 

conceptualizing this heterogeneity is Moffitt’s person-centered taxonomy, which groups youth 

into qualitatively distinct trajectories based on age of onset. Although a large body of research 

supports Moffitt’s distinction between life-course-persistent and adolescent-limited antisocial 

behavior, taxometric work has consistently indicated antisocial behavior to be a dimensional 

trait, without qualitative differences between persons. In other words, the theory implies that the 

field should use person-centered models, but the empirical data suggest antisocial behavior 

should be modeled using variable-centered statistical methods. Given this disconnect, it is 

surprising that nearly all longitudinal studies to date have employed only person-centered or 

variable-centered approaches, but not both. The present study sought to address this gap by 

examining the development of RB from middle childhood through to emerging adulthood via 

variable-centered and person-centered modeling, with socioeconomic disadvantage and peer 

delinquency considered as moderators. Both sets of analyses indicated that age-related 

increases in RB were normative, but variable-centered models were less useful for illuminating 

the association between socioeconomic disadvantage and the slope of RB. Across all analyses, 

peer delinquency emerged as a robust predictor of both persistence and escalation, consistent 

with Moffitt’s hypotheses regarding the role of delinquent peer affiliation in trajectories of ASB. 

Overall, findings provided some additional support for Moffitt’s taxonomy but also pointed to 

differences between trajectories that were a matter of degree, rather than kind, in line with the 

dimensional nature of youth behavioral problems. 

67 

 
 
Introduction 

Antisocial behavior (ASB) comprises a broad range of actions that violate societal norms 

and/or the rights of other people. Behaviors range from relatively minor rule violations (e.g., 

lying, cheating) to more serious offenses that violate the law (e.g., robbery, assault). One 

defining characteristic of ASB is its continuity within-persons across early development. That is, 

the youth who exhibit the most behavioral problems in childhood often follow a pattern of 

persistent offending, culminating in a series of poor adult outcomes (Odgers et al., 2008). 

Despite this continuity, mean levels of ASB are known to change dramatically over time, with 

ASB observed to increase as much as tenfold in frequency during adolescence (Moffitt, 1993). 

The adolescent-increase in ASB is largely attributable to a spike in non-aggressive rule-

breaking behaviors (RB), such as theft and vandalism, which are rare during childhood but 

increase in both prevalence and incidence during the teen years. These developmental patterns 

are consistent with the traditional conceptualization of the age-crime curve, one of the most 

robust findings in the field of criminology, with more youth observed to engage in delinquent 

activities during adolescence than at any other life stage (Hirschi & Gottfredson, 1983). 

In what is now a classic taxonomy, Moffitt (1993) proposed the peak of the age-crime 

curve to represent the activities of two qualitatively distinct groups of youth: life-course-

persistent offenders and adolescent-limited offenders. Life-course-persistent offenders were 

hypothesized to exhibit a broad spectrum of problem behaviors, often beginning with physical 

aggression (AGG) during the preschool years and escalating to include more severe aggressive 

and non-aggressive offenses throughout childhood and adolescence. Adolescent-limited 

offenders, by contrast, represented typically developing youth whose engagement in (primarily 

non-aggressive) ASB was transient and largely attributable to social factors, such as affiliation 

with delinquent peers. The remaining youth were thought to abstain from ASB throughout 

childhood and adolescence.  

One essential feature of Moffitt’s framework is its qualitative approach, with youth 

68 

 
classified into subgroups based on shared characteristics (e.g., age of onset). While youth 

within a particular subgroup are hypothesized to generally follow the same trajectory of ASB, 

trajectories are thought to differ between subgroups not only by degree but also by kind (e.g., 

escalating versus desisting). Each subgroup is also thought to be characterized by distinct 

correlates, with life-course-persistent ASB linked to more childhood risk factors and poorer life 

outcomes than all other trajectories. In Moffitt’s original framework, however, there was not a 

one-to-one correlation between the groups’ degree and kind of engagement in ASB and their 

psychological functioning. Specifically, Moffitt posited that youth who demonstrated adolescent-

limited ASB were in fact better adjusted than those who abstained completely, as the latter were 

believed to exhibit personality traits (e.g., behavioral inhibition, poor social skills) that left them 

excluded from adolescent peer networks (Moffitt, 1993).  

Since the original publication, Moffitt’s taxonomy has been updated in two key ways 

(Moffitt, 2003). First, the taxonomy expanded to include a childhood-limited group of youth, who 

largely desisted from ASB by adolescence. Second, the framework shifted to a more continuous 

conceptualization of the groups’ relative psychological functioning. That is, the abstainers were 

viewed as the healthiest group, followed by the childhood-limited, adolescent-limited, and life-

course-persistent groups, respectively. Follow-up data in adulthood from the Dunedin 

Longitudinal Study indicated that those youth who abstained from ASB had the best outcomes 

across domains, whereas adolescent-onset youth were second only to those with life-course-

persistent trajectories in psychological, financial, and legal problems (Moffitt, Caspi, Harrington, 

& Milne, 2002). In other words, adolescent-onset ASB was both relatively normative and linked 

to a series of poor outcomes in early adulthood.     

Consistent with the relatively normative nature of adolescent engagement in ASB, 

person-centered studies spanning adolescence have often reported a modal trajectory of 

increasing RB (e.g., Diamantopoulou, Verhulst, & van der Ende, 2011; Mata, 2013). Person-

centered studies have largely used a method developed by Nagin, termed semi-parametric 

69 

 
mixture modeling (Nagin, 1999). In this approach, participants are assigned to latent, or 

unobserved, trajectory groups based on their posterior probability of group membership (i.e., 

their likelihood of belonging to a particular group given their scores on the outcome measure at 

each timepoint). Although groups identified via semi-parametric mixture modeling are free to 

differ from one another by intercept (i.e., level of the outcome measure, often at baseline) and 

slope (i.e., rate of change), participants within the same group are assumed to share the same 

trajectory (see Figure 3.1a).  

Using these methods, Diamantopoulou and colleagues found that even “low” trajectory 

groups, which comprised the majority of participants, exhibited increases in RB between ages 

11 and 18. Likewise, nearly half of participants in the Montreal Longitudinal-Experimental Study 

followed a trajectory of moderate/increasing non-violent offending during adolescence, while a 

much smaller percentage of the sample (<10%) followed a chronic trajectory (Fontaine, 

Lacourse, Vitaro, & Tremblay, 2014). Consistent with Moffitt’s updated theory, both the 

“moderate” and “chronic” trajectory groups demonstrated poorer adult outcomes than did the 

“low” group. Other person-centered studies have similarly found many, if not most, youth to 

exhibit time-limited increases in RB but relatively few to demonstrate chronic engagement 

(Barker et al., 2007; van Lier et al., 2009), as Moffitt hypothesized. 

Despite the support of Moffitt’s theories in person-centered analyses, it is worth noting 

that some studies of ASB report variability within groups in baseline level and rate of change 

(e.g., Gardner, 2006; Schaeffer et al., 2003, 2006). These studies make use of growth mixture 

modeling (GMM), a more flexible alternative to Nagin’s approach that allows for both between 

and within-class variation (Muthén & Muthén, 2000). Although relatively few studies have 

leveraged GMM, these findings nevertheless reflect a deviation from Moffitt’s hypotheses. 

Consistent with this, other work has strongly indicated that person-centered methods do not 

accurately represent the underlying structure of ASB. Specifically, taxometric analyses have 

found data on youth ASB to conform to simulated dimensional data but not to simulated 

70 

 
categorical data, and these findings persisted to both cross-sectional and longitudinal analyses 

(Walters, 2011; Walters & Ruscio, 2013). Indeed, Burt (2012) argued that Moffitt’s taxonomy 

could be re-conceptualized in dimensional terms, with childhood-onset and adolescent-limited 

ASB linked to AGG and RB, respectively. Because engagement in ASB thus varies continuously 

across the population, some researchers argue that it has no qualitatively distinct underlying 

classes and is best-represented using variable-centered models (e.g., Walters & Ruscio, 2013).  

In longitudinal variable-centered approaches, growth curve models are used to quantify 

the mean change over time in the trait of interest across the entire sample (see Figure 3.1b; 

McArdle & Epstein, 1987). As in person-centered approaches, growth curve models capture 

individual patterns of continuity and change by the intercept and slope growth factors. Rather 

than considering the sample to comprise distinct subgroups, however, growth curve models 

treat all participants as belonging to the same underlying population, with patterns of growth that 

vary continuously from one another. The means of the intercept and slope represent the typical 

pattern of growth across the entire sample, whereas the variances of the intercept and slope 

represent individual differences in the baseline level of the outcome and the rate of change, 

respectively (Preacher, Wichman, MacCallum, & Briggs, 2008). As an example, a variable-

centered analysis of RB might find the mean linear slope to be positive, and the intercept and 

slope variances to each be significant. These findings would suggest that the average 

participant exhibited an increase in RB over time, but that there were significant between-person 

differences in both initial level and rate of change. 

In some ways, variable-centered studies of RB have yielded results consistent with 

those from the person-centered literature. For example, the sample-wide trend is typically one of 

increasing RB during adolescence (e.g., Doornwaard, Branje, Meeus, & ter Bogt, 2012; Padilla-

Walker, Memmott-Elison, & Coyne, 2018), followed by declining or plateauing engagement by 

early adulthood (e.g., Doornwaard et al., 2012). Moreover, most studies have reported 

significant variability between persons in both the intercept and slope of RB (Carroll, Mikhail, & 

71 

 
Burt, 2023), as well as significant differences in baseline engagement as a function of biological 

sex (e.g., Yan, Schoppe-Sullivan, & Beauchaine, 2021), neighborhood disadvantage (Ingoldsby, 

2003; Thaweekoon, 2006), and exposure to delinquent peers (Ingoldsby, 2003). 

Nevertheless, and despite their ability to correctly model the actual development of RB 

as a dimensional trait, variable-centered methods have limited potential to capture the 

heterogeneity observed in most samples. Even if the mean slope indicates a small age-related 

increase in RB, for example, a small percentage of participants may decline in their engagement 

over time, and others may exhibit a large increase. The primary strength of person-centered 

methods lies in their ability to identify these subgroups of youth, whose differing patterns of 

engagement cannot be fully captured in a single slope estimate. Indeed, proponents of person-

centered approaches contend that the groups serve as an approximation of participants’ 

trajectories and need not exist in reality to be theoretically and empirically useful (Nagin & 

Tremblay, 2005). For instance, person-centered analyses could facilitate the identification of risk 

factors that distinguish between youth who exhibit large adolescent-increases in RB and their 

counterparts whose engagement falls within normal limits, a finding with clear implications for 

intervention work. 

The considerations discussed thus far raise a key question for developmental 

researchers, namely, how can trajectories of ASB be modeled in a way that takes into account 

both developmental theory and the underlying structure of ASB data? That is, the most widely 

accepted developmental theory is Moffitt’s person-centered taxonomy, which provides a useful 

framework for understanding the heterogeneity within ASB but does not accurately represent its 

dimensional structure. One way to resolve this apparent disconnect is to leverage both person-

centered and variable-centered statistical approaches in the same dataset, which allows for a 

far more complete and accurate understanding of development than either can provide by itself. 

A recent review (Carroll, Mikhail, et al., 2023) began the process of integrating these two 

approaches through a conceptual review of the developmental literatures on antisocial behavior. 

72 

 
Of the 48 studies of RB included in the review, only one (Ingoldsby, 2003) leveraged both 

statistical techniques, with largely consistent results across methods. For instance, the modal 

trajectory was one of moderate RB that declined over time, and the growth curve model also 

indicated a mean-level decline. Both sets of analyses also identified exposure to neighborhood 

disadvantage and delinquent peers as predictors of RB trajectories. However, participants were 

only followed until age 10, meaning that neighborhood and peer influences were not examined 

as predictors of RB trajectories during adolescence. This represents a potentially important gap 

in the literature, in light of studies indicating that the extent of delinquent peer affiliation 

increases for most youth from childhood to adolescence (Wang & Dishion, 2012; Yoon, Yoon, 

Yoon, & Snyder, 2019). Neighborhood and peer influences may also be more salient to 

behavioral outcomes in adolescence relative to childhood, given age-related increases in 

mobility and independence (Leventhal & Brooks-Gunn, 2000). In short, because no study of RB 

to date has leveraged both person-centered and variable-centered approaches together in a 

sample spanning adolescence, the factors giving rise to the varied patterns of RB (i.e., 

desistance, persistence, and escalation) typically observed throughout early development are 

not well understood. 

Present Study 

The aim of the present study was to illuminate the development of RB from middle 

childhood into emerging adulthood in a large, population-based sample enriched for 

neighborhood disadvantage. Trajectories of RB were modeled via a series of variable-centered 

and person-centered analyses, with biological sex, race/ethnicity, socioeconomic disadvantage, 

and peer delinquency examined as potential moderators. In light of theoretical work that views 

youth development as embedded within multiple contexts (Bronfenbrenner, 1988) and empirical 

studies indicating that familial and neighborhood disadvantage each contribute uniquely to youth 

outcomes (Carroll, et al., 2023; Kupersmidt et al., 1995), we examined both household income 

and a composite measure of neighborhood deprivation as moderators. Consistent with prior 

73 

 
work (e.g., Doornwaard et al., 2012; Isen et al., 2022), we hypothesized that 1) the mean level 

of RB would increase across the study period, 2) there would be either three or four RB 

trajectory groups, and 3) male sex, lower household income, greater neighborhood 

disadvantage, and greater peer delinquency would predict higher levels of RB at baseline as 

well as persistent and escalating engagement across adolescence.  

Transparency and Openness 

Methods 

The present study was not preregistered. Because of the language in the informed 

consent document at intake, we cannot post the data publicly, but they can be obtained from the 

primary author upon reasonable request. We report how we determined our sample size, all 

data exclusions, all manipulations, and all measures in the study. All procedures were approved 

by the primary author’s institutional review board and are in compliance with the ethical 

standards of the relevant national and institutional committees on human experimentation and 

with the Helsinki Declaration. Child participants provided informed assent, and parents provided 

informed consent for themselves and their children.  

Participants 

The population-based Michigan State University Twin Registry (MSUTR) includes 

several independent twin projects. Participants in the current study were drawn from the Twin 

Study of Behavioral and Emotional Development in Children (TBED-C), a study within the 

MSUTR. The TBED-C includes a population-based arm (N=528 families), and an ‘under-

resourced’ arm for which inclusion criteria also specified that participating families lived in 

neighborhoods with neighborhood poverty levels at or above the Census mean at study onset 

(10.5%) (N=502 families). To recruit families at intake, the Department of Vital Records in the 

Michigan Department of Health and Human Services (MDHHS) identified twins in our age-range 

via the Michigan Twins Project (MTP), a population-based registry of more than 32,000 twins in 

Lower Michigan recruited via birth records. The Michigan Bureau of Integration, Information, and 

74 

 
Planning Services database was used to locate family addresses no more than 120 miles from 

East Lansing, Michigan through parent drivers’ license information. Pre-made recruitment 

packets were then mailed to parents by the MDHHS. Parents who did not respond to the first 

mailing were sent additional mailings roughly one month apart until either a reply was received 

or up to four letters had been mailed.   

This recruitment strategy for the TBED-C yielded an overall response rate of 57% for the 

under-resourced arm and 63% for the population-based arm. Other recruitment and sampling 

details can be found in prior publications (i.e., Burt & Klump, 2019). The two arms of the study 

were analyzed jointly for the current analyses and were 48.7% female and 51.3% male. Across 

the full sample, participants endorsed racial/ethnic identities in the following proportions: white 

(non-Latinx), 82%; Black, 10%; Latinx, 1%; Asian, 1%; Indigenous, 1%; multiracial, 6%. 

However, families in the under-resourced arm, but not the population-based arm, were more 

racially diverse than the local population (e.g., 14% Black and 77% white in the under-resourced 

arm versus 5% Black and 87% white in the population-based arm; 5% Black and 85% white in 

the local area Census).  

Because 90+% of TBED-C families were recruited out of the MTP, we were able to use 

the MTP data to compare families who chose to participate in TBED-C with those who were 

recruited but did not participate. TBED-C families were generally representative of recruited but 

non-participating families. As compared to non-participating twins, participating twins reported 

similar levels of conduct problems, emotional symptoms, and hyperactivity (d ranged from -.08 

to .01 in the population-based arm and .01 to .09 in the under-resourced arm; all ns). 

Participating families also did not differ from non-participating families in paternal felony 

convictions (d = -.01 and .13 for the population-based and under-resourced arms, respectively), 

rate of single-parent homes (d = .10 and -.01 for the population-based and under-resourced 

arms, respectively), paternal years of education (both d ≤ .12), or maternal and paternal alcohol 

problems (d ranged from .03 to .05 across the two arms). However, participating mothers in 

75 

 
both samples reported slightly more years of education (d = .17 and .26, both p < .05) than non-

participating mothers. Maternal felony convictions differed across participating and non-

participating families in the population-based arm (d = -.20; p < .05) but not in the under-

resourced arm (d = .02). In short, our recruitment procedures appear to have yielded a sample 

that was representative of both recruited families and the general population of the State of 

Michigan. 

Behavioral data regarding the children’s RB were collected at up to three time points. All 

2,060 twins in 1,030 families were assessed once in middle childhood (ages 6-11; Mage 

(SD)=8.02 (1.49)) as part of the TBED-C (Wave 1). TBED-C participants residing in modestly-to-

severely disadvantaged neighborhoods (N = 768 families) are currently being recruited for 

reassessment as adolescents up to two times, 18-months apart, through the Michigan Twin 

Neurogenetics Study (MTwiNS). For the present study, the sample sizes at Waves 2 and 3 were 

760 and 414, respectively.  

Procedure 

TBED-C families completed their intake assessments between 2008 and 2015. 

Assessment teams consisted of two research assistants and at least one paid staff member and 

took 4-5 hours to complete (lunch was provided). Families completed questionnaires, 

interviews, and videotaped interactions during the in-person assessment. The primary caregiver 

(nearly always the mother) participated alongside the twins. Assessments typically took place in 

our university laboratory (903 families). If families were unable or unwilling to travel, however, 

assessments took place in participants’ homes (13%). Families with older twins were less likely 

to complete home visits than on-site assessment visits (Cohen’s d=-.24, p=.01), as were 

families that identified as White (d=-.21; p<.05). Similarly, families with fewer financial resources 

were more likely to participate in home visits versus on-site assessments (d=-.32; p=.004). 

However, families completing on-site assessments versus home visits did not differ in maternal 

education, zygosity of the twins, or sex of the twins (ds ranged from -.02 to .16, all ns). The 

76 

 
MTwiNS follow-up assessments were also completed in-person. The twins and their primary 

caregiver completed an in-person assessment lasting 4-8 hours at either the East Lansing or 

Ann Arbor-based laboratories.  

Measures 

RB 

Youth RB was defined via a combination of caregiver, teacher, and youth reports. The 

primary caregiver (nearly always the twins’ mother) completed the 17-item RB scale from the 

Child Behavior Checklist (CBCL; Achenbach & Rescorla, 2001) at all three waves, rating the 

extent to which a series of statements described the child’s behavior over the past six months 

using a three-point scale (0=never to 2=often/mostly true). Caregiver-informant reports were 

available for 99% of participating twins at Wave 1, 99% at Wave 2, and 90% at Wave 3. 

Teacher reports of RB (12 items) were obtained at the first two assessment waves via the 

Achenbach Teacher Report Form (TRF; Achenbach & Rescorla, 2001). Teachers rated the 

twins’ behaviors during the preceding six months using the three-point scale described above. 

The teachers of 115 participants were not available for assessment (because the twins were 

home-schooled or because parental consents to contact the teachers were completed 

incorrectly, etc.). Our teacher participation rate across both subsamples was 86% at Wave 1 

and 60% at Wave 2, with teacher reports available for 1,551 and 453 participants, respectively. 

In addition, the twins completed the Youth Self-Report (YSR) at Waves 2 and 3, reporting on 

their own RB (15 items) during the preceding six months using the three-point scale described 

above (Achenbach & Rescorla, 2001). Reports were available from 99% and 83% of the twins 

at Waves 2 and 3, respectively. 

The RB scale from the Achenbach family of instruments assesses a wide range of 

behaviors, including minor rule violations (e.g., lying, cheating) as well as illegal behaviors (e.g., 

vandalism, setting fires). The latter had extremely low base rates in our sample, particularly at 

the childhood assessment. We thus constructed a more streamlined scale of RB items, 

77 

 
reflecting behaviors that were present to some degree at all assessment waves, via a series of 

item response theory (IRT) analyses. We also conducted a series of measurement invariance 

analyses to confirm that the items retained in the IRT models assessed RB consistently across 

measurement waves (see Supplementary Methods in Appendix C). 

Familial Socioeconomic Disadvantage 

Familial socioeconomic disadvantage was assessed via maternal reports of annual 

household income at all assessment waves. Household income was measured on a 10-point 

Likert scale (1 = <$10,000 to 10 = >$50,000) at Wave 1, and on a 13-point Likert scale (0 = 

<$5,000 to 12 = >$90,000) at Waves 2 and 3. To account for the different scales, we 

constructed a new eight-point scale (1 = <$10,000 to 8 = >$50,000) (e.g., families reporting an 

income <$5,000 at Wave 2 were coded as belonging to the <$10,000 bracket). Mean composite 

scores were computed to represent average income across all three waves. 

Neighborhood Socioeconomic Disadvantage 

Neighborhood disadvantage was defined via the Area Deprivation Index (ADI; Kind & 

Buckingham, 2018), which comprises 17 indices of Census block group disadvantage (e.g., 

poverty rate, income disparity). Data were weighted according to the factor loadings identified 

by Kind & Buckingham (2018), and weighted variables were summed to create a deprivation 

index score for each Census block group. Families were assigned a percentile indicating the 

level of deprivation in their block group relative to that of all U.S. block groups, with higher 

scores indicating greater deprivation. Mean ADI scores were 57.24 (SD=22.67; range: 2-99), 

61.57 (SD=20.94; range: 6-99), and 60.71 (SD=21.62; range: 12-99) at the TBED-C and 

MTwiNS assessments, respectively. These scores correspond to a moderate level of 

neighborhood disadvantage, albeit with considerable variation across participating families. At 

the TBED-C assessment, for instance, Census tract poverty rates ranged from 0 to 93%, with a 

mean of 19%. Mean ADI scores were computed to represent the average level of neighborhood 

deprivation over time. 

78 

 
Peer Delinquency 

Primary caregivers reported on their children’s peer group affiliations at Waves 1 and 2 

using the five-item Peer Delinquency scale from the Friends Inventory (Walden, McGue, Burt, & 

Elkins, 2004). Caregivers were instructed to provide ratings for each child’s entire peer group, 

with items scored on a four-point scale (1 = “none of my child’s friends are like that” to 4 = “all of 

my child’s friends are like that”; e.g., “my child’s friends break the rules”). Caregiver reports were 

available for 96.4% and >99% of participants at Waves 1 and 2, respectively. The twins also 

reported on their own peer groups’ delinquency at Waves 2 and 3 using the Friends Inventory. 

Items are nearly identical across parent and child versions of the Friends Inventory, with minor 

alterations in wording. Self-reports were available for >99% of participants at both waves. 

Parent and child reports were averaged across waves to represent mean peer delinquency over 

time. 

Race/Ethnicity 

Given inequalities in exposure to socioeconomic disadvantage across racial/ethnic 

identities (Peterson & Krivo, 2009), race was included as a covariate in all longitudinal analyses, 

coded dichotomously as white/racially marginalized due to the composition of our sample.  

Data Analytic Strategy 

Multi-informant mean RB scores were computed for each assessment wave based on 

the items retained in the final IRT model (see Supplementary Methods and Results in Appendix 

C for additional details, including confirmation of measurement invariance). Mean RB scores 

were log-transformed to adjust for positive skew and were subsequently examined as the 

outcome measures in the variable-centered and person-centered analyses. Continuous 

covariates (i.e., income, neighborhood disadvantage, and peer delinquency) were standardized 

for ease of interpretation. 

Variable-Centered Analyses 

Analyses were conducted in Mplus 8.4 (Muthén & Muthén, 1998-2019) using a robust 

79 

 
maximum likelihood estimator and full information maximum likelihood estimation (FIML) to 

account for missing data on the outcome. We examined phenotypic changes in the frequency of 

RB over time across the full sample via latent growth curve modeling, which quantifies within-

person changes in the outcome of interest as well as between-person variation in these 

changes (McArdle & Epstein, 1987). First, unconditional models were run to determine whether 

the frequency of RB increased, decreased, or remained constant with age. Age was examined 

in lieu of assessment wave as the metric of time given the wide range of ages included at each 

assessment (see Table S3.1). The definition variable approach was used to scale time in years, 

with the factor loadings constrained to vary by age (Sterba, 2014). Because the definition 

variables must be complete for the model to be estimated correctly, placeholder values were 

imputed to represent age for participants with missing data (Grimm, Ram, & Estabrook, 2017). 

Specifically, the mean number of years that elapsed between waves was added to participants’ 

ages at Wave 1 if data were missing at Waves 2 and/or 3. Age was available for all participants 

at Wave 1. Imputed values had no impact on estimation, as missing data on the definition 

variables was connected with missing data on the outcomes.  

Intercepts were centered to represent RB at age six, the youngest age at the first TBED-

C assessment wave. The slope factor represented annual change. Because participants were 

assessed on only three occasions, we did not examine non-linear growth. Initial analyses 

compared the respective fits of an intercept-only (“no-growth”) model and a model allowing for 

linear growth. The former had three parameters (intercept mean, intercept variance, and 

residual variance), and the latter had six (intercept and slope means, intercept and slope 

variances and their covariance, and residual variance). Model fit was evaluated with four 

likelihood-based indices: the chi-square difference test (Bollen, 1989), Akaike information 

criterion (AIC; Akaike, 1987), Bayesian information criterion (BIC; Raftery, 1995), and sample-

size adjusted Bayesian information criterion (SABIC; Sclove, 1987). For all indices, lower values 

indicated better model fit. In the unconditional models, the CLUSTER command was used to 

80 

 
account for the nesting of twins within families. 

The best-fitting model was selected for the conditional growth curve analyses. RB was fit 

conditional upon sex, race, income, ADI, and peer delinquency scores, all of which were 

averaged across waves and subsequently examined as time-invariant predictors of the intercept 

and slope. Because Mplus employs listwise deletion for missing covariates, all available data 

were analyzed in the conditional growth curve model by using multiple imputation (Little & 

Rubin, 2019), with five imputed datasets. As a result, estimates of the parameters and standard 

errors reflect the uncertainty that is due to missing data. In Mplus, it is not possible to include 

the CLUSTER command while using multiple imputation. Thus, in the conditional LGM, the data 

were structured to be in “family-wide” format, with parameters constrained to equality across 

both twins in a pair. The conditional LGM also accounted for the difference in within-pair 

resemblance for monozygotic (identical) and dizygotic (fraternal) twins, as identical twins are 

expected to resemble one another more so than fraternal twins on virtually every trait 

(Turkheimer, 2000).  

Person-Centered Analyses 

Latent class growth analyses were conducted in Mplus 8.4 using a robust maximum 

likelihood estimator and FIML to account for missing outcome data. We examined participants’ 

growth in RB as a function of their age at each assessment wave. We began with a one-class 

model and progressed through to a six-class model, consistent with the steps outlined by 

Muthén (2004) and with the approach employed in prior studies examining trajectories of ASB 

(e.g., Odgers et al., 2008). All models allowed the groups to differ from one another by intercept 

and slope but constrained within-group intercept and slope variances to zero. The optimal 

model was determined via consideration of two fit indices, the AIC and BIC, as well as the 

interpretability of the class solution (i.e., the percentage of the sample assigned to each class). 

Biological sex, race, income, ADI, and peer delinquency were subsequently examined as time-

invariant predictors of class membership, using five imputed datasets to account for missing 

81 

 
values. Notably, it is not possible to account for clustering when examining age as the metric of 

time in Mplus mixture models, nor was it feasible to examine the data in “family-wide” format 

because both twins in a pair would be assigned to the same trajectory by default. Thus, we did 

not account for clustering in our main analyses but ran supplemental analyses with one twin 

randomly selected per family, discussed in greater detail later in the manuscript. 

Descriptives 

Results 

Descriptive statistics and correlations across assessment waves are shown in Table 3.1, 

and descriptives by age are shown in Table S3.1. Participants evidenced moderate rank-order 

stability in RB over time. Paired-sample t tests indicated that mean levels of RB did not change 

significantly across assessment waves, although the increase from Wave 2 to Wave 3 was 

marginally significant (t(373) = 1.894, p = .059). At Waves 1 and 2, males had higher levels of 

RB than females (Cohen’s ds were .36 and .28, respectively, both ps <.001); sex differences 

were not significant at Wave 3. In addition, ADI and income were significantly correlated with RB 

at the first two time points, but not at the third. Peer delinquency was significantly correlated with 

RB at all waves. All correlations were in the expected direction.  

Variable-Centered 

Unconditional models 

Initial analyses compared the respective fits of an intercept-only model and a linear 

growth model to test the hypothesis that engagement in RB increased with age. The linear 

growth model fit better than the intercept-only model according to all fit indices (see Table 3.2) 

and was thus selected as the final unconditional model. At age six, the average RB score was 

.832 (SE = .029, p< .001), and scores increased by .018 points per year on average (SE = .005, 

p< .001), consistent with hypotheses. There was significant variation between-persons in both 

baseline level (b = .538, SE = .051, p< .001) and rate of change (b = .005, SE = .001, p< .001). 

The estimated covariance between the intercept and the slope was -.038 (SE = .006, p< .001), 

82 

 
indicating that participants with higher levels of RB at baseline exhibited somewhat less of an 

increase with age. 

Conditional model 

We subsequently examined ADI, income, peer delinquency, sex, and race as predictors 

of continuity and change in RB. As hypothesized, lower income, greater peer delinquency, and 

male sex predicted higher levels of RB at baseline. Income, peer delinquency, and sex also 

predicted rate of change (see Table 3.2). Specifically, males increased somewhat less in their 

RB over time relative to females, whereas participants with higher levels of peer delinquency 

exhibited a greater increase. Unexpectedly, higher income also predicted a greater increase in 

RB over time. Neither ADI nor race predicted the intercept or the slope. 

Person-Centered 

Unconditional models 

Initial analyses tested the hypothesis that there would be either three or four distinct RB 

trajectories. Model fit improved as the number of classes increased (see Table 3.3); however, 

the five and six-class models each contained groups that were not meaningfully distinct from 

one another. For example, the five-class model contained two groups with high, persistent RB. 

Based on both fit and theoretical considerations, the four-class model was selected as the 

optimal solution. As shown in Figure 3.2, the four-class model comprised two trajectories with 

low RB at baseline and two with high RB at baseline. The former were distinguished by the 

extent of their increase over time; one group, which comprised nearly 40% of the sample, 

exhibited a small increase with age, while the other, smaller group escalated substantially. Of 

the groups with high RB at baseline, one fully desisted by the end of adolescence, whereas the 

other followed a chronic trajectory. 

Conditional model 

We first examined the predictors of membership in the chronic trajectory relative to the 

normative group. These groups differed as a function of ADI, income, peer delinquency, and 

83 

 
sex. Specifically, greater neighborhood deprivation, lower income, greater affiliation with 

delinquent peers, and male sex predicted chronic RB trajectories. To elucidate the factors 

predicting the slope, we subsequently examined each pair of groups with similar RB at baseline. 

ADI and peer delinquency distinguished between the escalating and normative groups, such 

that youth who resided in disadvantaged neighborhoods and associated with more delinquent 

peers were more likely to increase substantially in their RB. High levels of peer delinquency also 

predicted membership in the chronic trajectory relative to the desisting trajectory, as did low 

income. Neither sex nor race distinguished between either pair of trajectory groups (see Tables 

3.3 and S3.6 for conditional model results and Table 3.4 for trajectory group characteristics).   

Interpretive Integration 

Variable-centered and person-centered analyses both indicated that an increase in RB 

over time was normative. Moreover, lower household income, greater peer delinquency, and 

male sex all predicted elevated RB at baseline, as they predicted the intercept of the growth 

curve and distinguished between the chronic and normative trajectory groups. Regarding 

predictors of the slope, however, results were consistent across modeling approaches for some 

but not others. Both sets of analyses identified peer delinquency as a critical predictor of 

escalating engagement. That is, greater peer delinquency predicted a greater age-related 

increase in RB in the growth curve model, as well as membership in the escalating and chronic 

groups relative to the normative and desisting groups, respectively, in the mixture models. 

However, familial affluence predicted a greater increase in RB in the growth curve, a finding that 

may reflect a statistical artifact, as those from wealthier families had lower levels of RB at 

baseline and thus had more room to increase over time. Person-centered results, by contrast, 

indicated that income distinguished between the chronic and desisting groups, and in the 

expected direction. Neighborhood deprivation also distinguished between the escalating and 

normative trajectories but did not predict the slope of the growth curve, possibly because the 

escalating trajectory comprised a small portion of the sample. Such findings suggest that 

84 

 
variable-centered approaches are limited in their ability to elucidate the factors predicting 

membership in small, yet clinically meaningful, trajectory groups. 

Discussion 

The present study aimed to elucidate trajectories of RB across early development by 

integrating results from person-centered and variable-centered modeling approaches. 

Consistent with our hypotheses, age-related increases in RB were normative. That is, the 

sample-wide pattern identified in the variable-centered analyses was one of increasing RB 

across the study period, and person-centered analyses yielded a modal trajectory characterized 

by a small increase. Taken together, these findings are consistent with the broader empirical 

literature, which has reported that most (but not all) youth follow trajectories of increasing RB 

during adolescence (e.g., de Haan, Prinzie, & Dekovic, 2012; Diamantopoulou et al., 2011). The 

person-centered results are also largely consistent with Moffitt’s taxonomy (Moffitt, 1993, 2003), 

with a subgroup of youth exhibiting a life-course-persistent trajectory and a larger portion of the 

sample demonstrating onset in adolescence. 

Nevertheless, the development of RB was far from uniform across individuals and 

trajectory groups. For instance, nearly half (46.6%) of the sample followed a trajectory 

consistent with Moffitt’s adolescent-limited class, with little-to-no engagement during childhood, 

but the two adolescent-limited groups identified here differed sharply in the magnitude of their 

age-related increase. The larger of the two groups exhibited a modest increase, consistent with 

the mean pattern identified in the growth curve, whereas the other group, which comprised 

15.5% of adolescent-increasers, eventually surpassed the chronic trajectory in level of 

engagement. The heterogeneous patterns of change identified here raise a question that 

person-centered approaches are uniquely positioned to answer, namely, what factors predict 

escalating RB among youth with few behavioral problems in childhood? 

The mixture models found ADI and peer delinquency to predict membership in the 

escalating trajectory relative to the normative trajectory. Specifically, those who escalated were 

85 

 
more likely to reside in disadvantaged neighborhoods and affiliate with delinquent peers, 

consistent with theory (Bronfenbrenner, 1988; Moffitt, 1993, 2003) and prior research (Carroll et 

al., 2023; Ettekal & Ladd, 2015). Higher levels of peer delinquency also predicted a greater age-

related increase in RB in the growth curve model. Results were not consistent across modeling 

approaches for ADI, however, as ADI did not predict the slope of the growth curve. These 

discrepant findings may be due to the size of the escalating trajectory, which comprised only 

~7% of participants and was likely not well-characterized via analyses that collapsed across the 

entire sample. Person-centered models were thus better positioned to illuminate the role of 

neighborhood disadvantage in predicting escalating RB, although both sets of analyses 

implicated peer delinquency as an important predictor of the slope.   

Integration of the person-centered and variable-centered results also yields a fuller 

understanding of the factors predicting desistance from RB. Lower levels of peer delinquency 

predicted membership in the desisting trajectory relative to the chronic one, consistent with the 

results for peer delinquency in the growth curve model. For income, however, the growth curve 

indicated that affluence was associated with a greater increase in RB with age, a finding that 

may reflect regression to the mean, as suggested by prior variable-centered simulations 

examining predictors of change (Marsh & Hau, 2002). In the person-centered analyses, by 

contrast, higher income predicted desistance relative to persistence. Thus, while the growth 

curve model accurately identified the association between income and RB at baseline, 

consideration of results from the mixture model was necessary to understand the association 

between income and rate of change. 

The present study also found that ADI and income incremented one another as 

predictors of membership in the chronic trajectory group relative to the normative one, 

consistent with prior work indicating that exposure to disadvantage in multiple contexts is 

particularly salient to the development of externalizing (Carroll et al., 2023; Kupersmidt et al., 

1995). Not only did disadvantage predict elevated RB at baseline in our sample in both sets of 

86 

 
analyses, but it also predicted escalation relative to normative engagement and persistence 

relative to desistance in the mixture models. These findings underscore the long-lasting 

implications of socioeconomic disadvantage for youth mental health. Youth who reside in 

disadvantaged neighborhoods may face economic insecurity, household overcrowding, and/or 

community violence, among other risk factors (Singh, 2003). Additional research is needed to 

determine which specific aspects of neighborhood disadvantage serve as ‘active ingredients’ for 

trajectories of persistent or escalating RB. In turn, our finding that familial affluence predicted 

desistance may partially reflect treatment effects, as youth from wealthier backgrounds who 

demonstrate childhood-onset RB may be more likely to access treatment than their less affluent 

peers. Future studies should examine whether access to psychological treatment explains the 

association between household income and recovery from RB. 

The variable-centered and person-centered results together also shed light on sex 

differences in RB. The former indicated that males exhibited greater engagement at baseline, 

consistent with the broader variable-centered literature on RB, which has typically found sex to 

predict the intercept (Carroll, Mikhail, et al., 2023). Male sex also predicted a somewhat smaller 

increase with age in the growth curve, a finding that is less consistent with prior research. 

Results from the mixture models help to clarify the association between sex and rate of change. 

Male sex predicted membership in the chronic trajectory relative to both the normative and 

escalating trajectories and also predicted desistance relative to escalation. In other words, 

males were somewhat overrepresented in the only group that exhibited a significant age-related 

decline, and this was reflected in the growth curve model. Notably, the chronic trajectory was 

nearly two-thirds male, consistent with Moffitt’s conceptualization of a predominantly male life-

course-persistent group.  

Limitations 

Our results should be interpreted in light of several limitations. First, participants were 

predominantly white. Although our sample was socioeconomically diverse and representative of 

87 

 
the racial/ethnic makeup of the State of Michigan, future studies are needed to determine 

whether the present findings generalize to youth from more diverse racial/ethnic backgrounds. 

Next, participants were assessed on up to three occasions, meaning it was not possible to 

model non-linear fluctuations in RB. This may be particularly salient during adolescence, when 

engagement in RB is often observed to peak and subsequently decline (e.g., Doornwaard et al., 

2012). Rather than follow a linearly increasing trajectory, for example, youth in the escalating 

group may have fluctuated in their engagement across adolescence. It is also unclear when, or 

if, youth in the escalating and chronic trajectories began to desist from RB. Results from prior 

studies that have followed participants into young adulthood have yielded mixed results as to 

when desistance occurs. In the Pittsburgh Youth Study, for instance, all trajectory groups 

exhibited near-zero levels of RB by the end of the study period (i.e., age 24) (Lacourse et al., 

2008), whereas in a cohort from the Dunedin Multidisciplinary Health and Development Study, 

two of four trajectories (comprising nearly 1/3 of the sample) continued to engage in ASB at age 

26 (Odgers et al., 2008). Subsequent studies with many assessment waves are needed to 

determine the typical timing of desistance from RB, as well as elucidate the factors predicting 

continued engagement into adulthood. 

Next, we modeled RB based on reports from multiple informants at each wave. Although 

the use of multiple informants is likely to yield a fuller picture of RB than any single informant 

could provide (Achenbach, McConaughy, & Howell, 1987), youth self-reports were not available 

at Wave 1, and teacher reports were not available at Wave 3 (i.e., only caregiver reports were 

obtained at all waves). In addition to confirming measurement invariance, we conducted a 

series of sensitivity analyses using caregiver-reported RB as the outcome measure. As shown 

in Tables S3.7 and S3.8, as well as Figure S3.1, results were largely consistent with those from 

the main analyses. Results differed somewhat with respect to the frequency of engagement in 

RB during adolescence, as the unconditional growth curve indicated a slight decrease in RB 

with age, and a smaller percentage of participants followed an escalating trajectory than in the 

88 

 
main analyses. These discrepancies likely reflect caregivers’ incomplete awareness of their 

children’s engagement in RB, as youth are both motivated to conceal these behaviors and 

increasingly able to do so as they get older (Burt, 2012). In support of this, meta-analytic work 

has found cross-informant correlations for a variety of psychiatric symptoms to be lower during 

adolescence than childhood (Achenbach et al., 1987). Nevertheless, the overall pattern of 

results was consistent, and peer delinquency remained a robust predictor of trajectory group 

membership. 

Another limitation concerns the nested structure of the data. As noted in the Methods, 

we could not include the CLUSTER command in the conditional LGM or in any of the mixture 

models. We accounted for clustering in the conditional LGM by examining the data in “family-

wide” format. As a verification, we also ran the conditional LGM with the data in “long” format, 

without controlling for clustering. As shown in Table S3.9, results were consistent with those 

from the main analyses. Notably, ADI predicted the intercept (in the expected direction), 

whereas it was marginally significant in the main analyses. For the mixture models, we ran the 

unconditional and conditional models with one twin per family randomly selected (see Table 

S3.10). The four-group solution was nearly identical to the solution from the full sample, with a 

modal trajectory of low/increasing RB, an escalating group, and two other groups distinguished 

by degree of desistance. Results were largely consistent in the conditional analyses as well, 

with elevated peer delinquency predicting membership in the escalating and chronic groups 

relative to the normative and desisting groups, respectively. Overall, we conclude that our 

results are not biased by the limitations of the software in accounting for clustering. 

A methodological limitation specific to the mixture models was that the average posterior 

probability for the escalating trajectory was low. In mixture modeling, each participant is 

assigned to the group for which their posterior probability of membership is highest. Average 

posterior probabilities are thus a measure of classification accuracy, as higher values indicate 

greater certainty in class assignment. The average probability for the escalating group (.29 in 

89 

 
the unconditional model and .59 in the conditional model) was likely related to the study’s 

accelerated longitudinal design, with incomplete follow-up at the second and third waves of data 

collection. When the sample was restricted to those with complete data at all waves (N=377), 

the posterior probability for the escalating group was .73, and the probabilities for all other 

groups were > .80, above the recommended threshold of .70 (Nagin, 2005). The class solution 

was fully consistent with that observed in the full sample. In addition, average posterior 

probabilities for all groups but the escalating trajectory were > .70 in the full sample, suggesting 

that participants were assigned to groups with a reasonable degree of confidence, despite the 

planned missingness at later waves. 

Lastly, although reports of neighborhood disadvantage, income, and peer delinquency 

were averaged across waves to represent each construct throughout the study period, they 

were considered as time-invariant predictors. Notably, the mean level of peer delinquency did 

not change significantly over time (b = -.001, p = .271), (although at Wave 2, youth reported 

significantly higher levels than caregivers did (t(748) = 6.26, p < .001), consistent with prior 

research on informant effects (Achenbach et al., 1987)). Although beyond the scope of this 

study, additional research is needed to elucidate the role of familial and neighborhood 

disadvantage and peer delinquency as time-varying predictors. In particular, studies should 

examine trajectories of peer delinquency and trajectories of RB in tandem to better understand 

the (likely reciprocal) relationship between the two. 

Implications 

Nevertheless, the present study has several important implications. First, the use of 

variable-centered and person-centered modeling approaches together yields a far more 

complete picture of development than either approach can provide by itself. Variable-centered 

methods are integral to modeling development across the full sample, which illuminates 

normative patterns and in so doing facilitates understanding of atypical development. In the 

present study, for example, the mean pattern in the growth curve indicated that a modest age-

90 

 
related increase in RB was normative, but that a sharp escalation was not. Variable-centered 

analyses were also useful for identifying predictors of elevated RB at baseline but were 

inconsistent in their ability to accurately model predictors of the slope. That is, some parameter 

estimates appeared to capture the actual association between predictor and slope (i.e., greater 

peer delinquency predicted a greater increase in RB), whereas others reflected regression to 

the mean artifacts (i.e., higher income predicted a greater increase in RB). Consideration of the 

person-centered results was necessary to determine which estimates from the growth curve 

were likely accurate and which were artifacts. Moreover, by allowing for the direct comparison of 

distinct trajectory groups, the person-centered analyses identified several factors predicting 

persistent and/or escalating engagement (i.e., low income, neighborhood deprivation, and peer 

delinquency), even though the sample likely did not comprise four qualitatively distinct 

trajectories (Walters & Ruscio, 2013). In short, both sets of analyses were needed to model the 

development of RB as a dimensional, yet highly heterogeneous, trait. 

Second, the present study implicates peer delinquency as a critical risk factor for the 

development of RB. Across all analyses, peer delinquency was the most robust and consistent 

predictor, as it predicted both the intercept and slope of the growth curve (in the expected 

direction) and distinguished between five of the six group pairings from the mixture models. That 

greater peer delinquency predicted escalation relative to desistance suggests that its effects are 

particularly salient during adolescence, consistent with Moffitt’s framework, which emphasizes 

the role of delinquent peer affiliation in the development of adolescent-limited ASB (Moffitt, 

1993, 2003). Although the association between peer delinquency and RB is well-established 

(e.g., Ettekal & Ladd, 2015; Mann et al., 2015), no study to date has leveraged variable-

centered and person-centered analyses to examine trajectories of peer delinquency and RB 

simultaneously. Increases in peer delinquency, for example, may predict increasing RB, 

evidenced by positive associations with the slope of RB and significant overlap with escalating 

RB trajectories. Given the unique strengths of each statistical approach, future studies that 

91 

 
make use of both are needed to elucidate the role of peer delinquency over time, which may 

represent a promising target for prevention and intervention efforts addressing adolescent RB.  

Overall, findings from both sets of analyses provide some additional support for Moffitt’s 

framework, given the sample-wide and modal patterns of increasing RB reported here and the 

consistent evidence in support of peer influences. Moreover, the person-centered findings 

indicated that some participants followed a chronic, or life-course-persistent, trajectory, but that 

a larger percentage demonstrated onset in adolescence, consistent with Moffitt’s original 

taxonomy (Moffitt, 1993). Despite these confirmations of the taxonomy’s hypotheses, the 

present findings diverge from theory in two key ways. First, age-related increases in RB were far 

from universal, with a substantial portion of the sample demonstrating desistance. That is, of the 

1,099 youth observed to exhibit childhood-onset RB, 694 (63%) recovered so completely that 

they engaged in less RB than even the normative group by the end of the study. Such findings 

suggest that recovery from early-onset ASB may be even more common than previously 

suggested (e.g., Moffitt et al., 2002). This raises questions about the nature of “life-course 

persistent” ASB. Second, the escalating group demonstrated such a dramatic increase that its 

engagement eventually surpassed that of the chronic group. This runs counter to Moffitt’s 

conceptualization of adolescent-limited and life-course-persistent ASB as qualitatively distinct 

syndromes with marked differences in severity. The present findings instead point to differences 

between the two trajectories that are a matter of degree and suggest that some youth with late-

onset ASB may become (at least) as impaired as those whose engagement begins in childhood. 

At the same time, some adolescent-limited youth likely conform to the pattern hypothesized by 

Moffitt, outgrowing their behavioral problems and successfully transitioning into adult roles 

(Moffitt, 1993). The likely distinctions within the classes that Moffitt proposed underscore the 

dimensional nature of RB (and ASB in general), as well as the need for both person-centered 

and variable-centered approaches to fully and accurately model trajectories across the early 

developmental period.  

92 

 
Table 3.1. Descriptive statistics and correlations 

Tables 

1.  ADIa 
2.  Incomea 
3.  Peer 

Delinquencyb 

4.  RB T1c 
5.  RB T2c 
6.  RB T3c 

1. 
- 
-.40 
.15 

.17 
.14 
.06 

2. 

- 
-.11 

-.24 
-.13 
-.09 

3. 

- 

.32 
.46 
.35 

4. 

5. 

6. 

- 
.31 
.30 

- 
.62 

- 

Mean (SD) 

6.91 
(1.77) 
1-8 
2030 
Note: Bold font indicates p<.05. aMean scores were computed based on reports at all waves. 

56.60 
(22.15) 
2-99 
2031 

2.74 
(3.08) 
0-23 
769 

6.32 
(1.25) 
5-13 
2021 

2.54 
(3.44) 
0-21 
2055 

2.91 
(3.26) 
0-21 
379 

Range 
N 

bMean scores were computed based on maternal and youth reports at all waves. cMulti-

informant mean scores were computed based on maternal and teacher reports at Wave 1, 

maternal, teacher, and youth reports at Wave 2, and maternal and youth reports at Wave 3. 

Scores were multiplied by 24 to correspond to the maximum number of items at any wave. 

93 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 3.2. Latent growth curve model fit statistics and parameter estimates 

Model 
Linear growth 
Means model 

Unconditional LGM model fit statistics 

-2lnL 
7767.02 
7834.63 

χ2 (df) 
- 

AIC 
7779.02 
67.61† (3)  7840.63 

BIC 
7812.81 
7857.52 

SABIC 
7793.74 
7847.99 

Parameter 

Mean 
Variance 
ADI  
Income  
Peer Delinquency  
Sex  
Race  

Mean 
Variance 
ADI 
Income  
Peer Delinquency 
Sex  
Race  

Intercept with 
slope 
Intercept 1 with 
Intercept 2 
          MZ 
          DZ 
Slope 1 with Slope 2 
          MZ 
          DZ 
Intercept 1 with 
Slope 2 
            MZ 
            DZ 
Residual variance 

Conditional LGM parameter estimates 
Estimate 

p-value 

S.E. 
Intercept 
.077 
.042 
.030 
.032 
.027 
.052 
.080 
Slope 
.012 
.001 
.005 
.005 
.005 
.008 
.012 
Covariances 
.005 

.040 
.041 

.001 
.001 

.005 
.005 
.021 

<.001 
<.001 
.062 
<.001 
<.001 
<.001 
.460 

.010 
<.001 
.713 
.016 
<.001 
.005 
.878 

<.001 

<.001 
<.001 

<.001 
<.001 

<.001 
<.001 
<.001 

.644 
.517 
.057 
-.162 
.170 
.268 
.059 

.030 
.006 
.002 
.012 
.018 
-.023 
-.002 

-.046 

.564 
.365 

.007 
.005 

-.054 
-.035 
.290 

Note: Multi-informant composite mean RB scores are the outcome measure. Bold font indicates 

p< .05. †Significant change in chi-square at p<.05.  

94 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 3.3. LCGA fit statistics and predictors of group membership 

Unconditional LCGA model fit statistics 

AIC 

BIC 

Smallest group 
(% of sample) 

# of groups 
1 
2 
3 
4 
5 
6 

7965.11 
7652.33 
7547.96 
7417.13 
7361.79 
7296.34 

7982.00 
7686.11 
7598.63 
7484.69 
7446.24 
7397.67 

100 
38.9 
6.0 
7.2 
4.4 
2.3 

Conditional LCGA parameter estimates from 4-group model 
Chronic v. normative 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

Estimate 
.453 
-.634 
2.072 

.868 
.327 

S.E. 
.135 
.126 
.201 

.295 
.314 

Escalating v. normative 
S.E. 
Estimate 
.154 
.460 
.147 
-.159 
.207 
1.762 

-.277 
.271 
Chronic v. desisting 

.284 
.357 

Estimate 
.177 
-.357 
1.212 

.582 
.090 

S.E. 
.140 
.119 
.162 

.301 
.311 

p-value 
.001 
<.001 
<.001 

.003 
.297 

p-value 
.003 
.279 
<.001 

.330 
.448 

p-value 
.205 
.003 
<.001 

.053 
.771 

Note: Multi-informant composite mean RB scores are the outcome measure. Bold font indicates 

p< .05. 

95 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table 3.4. Characteristics of RB trajectory groups 

N 
% male 

ADI 

Income 

Peer Delinquency 

RB T1 

RB T2 

RB T3 

RB T1 (log-
transformed) 
RB T2 (log-
transformed) 
RB T3 (log-
transformed) 

Class 1: 
normative 
811 
42.5 

51.72 (21.39) 
[4-99] 
7.23 (1.50) 
[1-8] 
5.88 (1.04) 
[5-11] 
.06 (.32) 
[0-3] 
.69 (1.00) 
[0-7] 
.77 (1.03) 
[0-4.50] 
.04 (.18) 
[0-1.39] 
.39 (.49) 
[0-2.08] 
.42 (.53) 
[0-1.70] 

Class 2: 
escalating 
149 
47.7 

Means (SD) 
[Range] 
64.03 (20.34) 
[21-99] 
7.05 (1.47) 
[1.33-8] 
6.57 (1.22) 
 [5-13] 
.21 (.52) 
[0-1.50] 
4.10 (2.93) 
[0-23] 
4.67 (3.39) 
[1.50-21] 
.13 (.32) 
[0-.92] 
1.49 (.53) 
[0-3.18] 
1.60 (.50) 
[.92-3.09] 

Class 3: 
desisting 
694 
54.8 

Class 4: 
chronic  
405 
64.4 

57.22 (22.30) 
[4-99] 
6.89 (1.79) 
[1-8] 
6.32 (1.17) 
[5-13] 
2.95 (1.76) 
[0-15] 
1.51 (1.46) 
[0-13.50] 
1.74 (1.52) 
[0-7.50] 
1.30 (.38) 
[0-2.77] 
.78 (.54) 
[0-2.67] 
.84 (.60) 
[0-2.14] 

62.42 (21.78) 
[2-99] 
6.28 (2.15) 
[1-8] 
7.11 (1.35) 
[5-11] 
7.67 (3.72) 
[1.50-21] 
6.37 (3.29) 
[1-16.50] 
6.42 (3.72) 
[1.50-18] 
2.07 (.43) 
[.92-3.09] 
1.91 (.43) 
[.69-2.86] 
1.89 (.47) 
[.92-2.94] 

Note: Multi-informant composite mean RB scores are the outcome measure.  

96 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5

4

3

2

1

0

5

4

3

2

1

0

Figures 

Person-centered: mixture model

8

9

10

11

12

13

14

15

16

17

Age (years)

Group 1: low, slight increase

Group 2: escalating

Group 3: desisting (childhood-limited)

Group 4: chronic

Variable-centered: growth curve model

8

9

10

11

12

13

14

15

16

17

Age (years)

Participant 1

Participant 2

Participant 3

Participant 4

Sample mean

Participant 5

i

g
n
k
a
e
r
B
-
e
u
R

l

i

g
n
k
a
e
r
B
-
e
u
R

l

a.

b.

Figure 3.1. Person-centered (a) and variable-centered (b) approaches to modeling the 

development of RB in a hypothetical sample. The person-centered model extracted four distinct 

trajectory groups, whereas the variable-centered model identified a mean-level increase across 

the entire sample. 

97 

 
 
 
 
i

g
n
k
a
e
r
B
-
e
u
R

l

2.5

2

1.5

1

0.5

0

Group 1: normative
(39.39%)

Group 2: escalating
(7.24%)

Group 3: desisting
(33.71%)

Group 4: chronic
(19.67%)

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Age

Figure 3.2. Unconditional four-group model of RB trajectories based on multi-informant reports. 

98 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
GENERAL DISCUSSION 

By leveraging multiple statistical approaches in a sample spanning nearly all of early 

development, the present studies substantially advanced our understanding of the origins of 

ASB. Study 1 found that more than one-third of the genetic and nonshared environmental 

contributions, respectively, to change in ASB were already present at the baseline assessment. 

Given the broad age range under study (i.e., ages three to 21 years), such findings indicate a 

remarkable degree of continuity over time in both genetic and environmental influences. 

Furthermore, the present findings align with prior research on youth psychopathology indicating 

that the most potent environmental influences are those that make children in the same family 

less alike (Plomin & Daniels, 1987). Study 1 clearly implicates (some of) these influences as 

enduring and systematic, rather than merely transient, and proposes several possible 

nonshared environmental factors that may systematically shape the development of ASB, 

including differential parenting, delinquent peer affiliation, and aspects of the family environment 

that are objectively shared but impact siblings in distinct ways (e.g., parental divorce; Goldsmith, 

1993). 

Although the analyses were not genetically informed, Studies 2 and 3 illuminated several 

putatively environmental contributions to trajectories of AGG and RB through their use of both 

variable-centered and person-centered methods. In Study 2, parent-child conflict predicted not 

only baseline engagement in AGG, but also change over time, such that youth experiencing 

less conflict with their mothers were far more likely to exhibit desistance. Put differently, low 

levels of conflict appeared to interrupt trajectories of elevated AGG, consistent with prior work 

identifying parent-child conflict as a key psychosocial risk factor for externalizing (Burt, Krueger, 

McGue, & Iacono, 2003). The prospect of differential parenting as a systematic, nonshared 

environmental influence on ASB trajectories is also bolstered by studies leveraging quasi-

experimental designs, which have found twin differences in parent-child conflict (Burt, McGue, 

Iacono, & Krueger, 2006) and harsh parenting (Burt et al., 2021) to predict twin differences in 

99 

 
ASB. Such findings support an association between parenting and ASB that is at least partially 

environmental in origin and may well be causal to some extent. Study 2 builds on this prior work 

by implicating the absence of conflict as a protective factor for youth who initially demonstrate 

high levels of AGG, although additional research leveraging genetically informative designs is 

needed to clarify whether the association between low conflict and desistance is also 

environmentally mediated. 

In turn, Study 3 found both neighborhood disadvantage and affiliation with delinquent 

peers to predict RB trajectories, consistent with prior research on the effects of the broader 

neighborhood (e.g., Lacourse et al., 2008) and peer (e.g., Mann et al., 2015) contexts. 

Moreover, the effects of neighborhood disadvantage were observed when controlling for 

household income, indicating a unique role of contextual disadvantage in the development of 

RB (consistent with findings from cross-sectional studies that disambiguated neighborhood and 

familial disadvantage (Carroll et al., 2023; Kupersmidt et al., 1995)). Although objectively shared 

by youth residing in the same home, the neighborhood context may impact each child in 

idiosyncratic ways. Neighborhood disadvantage is also a multifaceted construct (Singh, 2003), 

encompassing risk factors such as pervasive poverty, household overcrowding, and community 

violence, which may be shared by siblings to differing degrees. Differences between siblings in 

exposure to community violence, for instance, might predict differences in their engagement in 

RB. The association between peer delinquency and RB, on the other hand, may be more 

complex, given that youth are able to select into peer groups in a way that they do not typically 

select into neighborhoods. Studies leveraging cross-lagged designs have sought to clarify 

whether the association is more indicative of selection (i.e., youth who engage in ASB seek out 

delinquent peers) or socialization (i.e., peer delinquency confers risk for ASB). Although some 

work has pointed solely to a selection effect (Burt, McGue, & Iacono, 2009), most studies have 

found evidence of both selection and socialization (Dugré, Giguére, & Potvin, 2024; Kendler, 

Jacobson, Myers, & Eaves, 2008; Samek et al., 2016; Schwartz, Solomon, & Valgardson, 

100 

 
2019). Other work has indicated that the respective roles of selection and socialization are age-

dependent, with greater evidence of selection effects during adolescence relative to childhood 

(Kendler et al., 2008) as youth are increasingly able to select into environments that are 

consistent with their predispositions (Beam & Turkheimer, 2013). Additional longitudinal studies, 

including those incorporating genetically informative designs, are needed to elucidate the 

temporal associations between peer delinquency and RB (and ASB in general) throughout the 

early developmental period. 

 Some of the predictors discussed above may not be specific to either AGG or RB but 

rather confer risk for trajectories of ASB in general. Indeed, Burt and colleagues found parent-

child conflict to underlie comorbidity for multiple forms of externalizing (Burt et al., 2003), and 

Dugré et al. (2024) reported bidirectional associations between peer delinquency and AGG and 

RB, respectively. Furthermore, AGG and RB often co-occur, with correlations of moderate-to-

large magnitude typically reported between the two (rs .4-.6; Burt, 2013). Such findings raise the 

possibility that the high-risk trajectories identified via mixture modeling in Studies 2 and 3 largely 

comprise the same participants. Cross-tabulation results, shown in Table 4.1, help to clarify the 

extent of the overlap between trajectories of AGG and trajectories of RB in the present sample. 

As expected, participants evidenced some overlap in their group memberships; most 

participants who were assigned to the normative trajectory for one dimension of ASB, for 

instance, were also assigned to the normative trajectory for the other. In addition, nearly two-

thirds of those following a persistent AGG trajectory also exhibited persistent RB. However, the 

reverse was not true, as most participants who exhibited persistent RB followed either 

normative or desisting trajectories of AGG. In addition, most youth who followed desisting 

trajectories of one form of ASB did not do so for the other. Lastly, the majority of those following 

a trajectory of escalating RB, for which there was not an analogous trajectory of AGG, fell into 

the normative AGG group.  

These findings collectively indicate that associations between AGG and RB trajectories 

101 

 
are far from unity, consistent with results from the few extant studies examining overlapping 

trajectories (e.g., Mata & van Dulmen, 2012; Maughan et al., 2000). Far more work is needed to 

understand how AGG and RB develop in tandem as related yet distinct traits. Studies should 

leverage parallel process growth curve models to examine the development of AGG and RB, 

respectively, across entire samples, as well as parallel process mixture models to capture 

patterns of growth for distinct subgroups. As discussed above, it is also unclear which 

risk/protective factors are specific to each dimension of ASB and which confer, or reduce, risk 

for ASB in general. Of the 1078 youth following normative trajectories of AGG, for instance, why 

do 445 exhibit non-normative trajectories of RB? Moreover, how do these youth differ from 

those following persistent trajectories of both AGG and RB? Future research could answer 

these questions by examining predictors of joint trajectories via mixture modeling or via 

analyses that control for one dimension of ASB while examining engagement in the other.  

There are several other limitations to the present studies. First, all studies focused on 

changes in the frequency, rather than severity, of ASB, and some of the behaviors assessed 

were of relatively low severity (e.g., hot temper, breaks rules). Among those youth engaging in 

frequent and persistent ASB, severity typically increases with age (e.g., Copeland, Miller-

Johnson, Keeler, Angold, & Costello, 2007). Additional research is needed to elucidate when 

and how escalations in severity occur, as well as which specific behaviors comprise the highest-

risk trajectories. In addition, because comorbidity is associated with greater severity (e.g., 

Kessler, Chiu, Demler, & Walters, 2005), future studies should examine patterns of comorbid 

AGG and RB in conjunction with age-related changes in the specific types of offenses 

committed. 

Next, it is unclear to what extent the effects of the risk and protective factors under study 

differed across sex. The present studies found males to exhibit higher levels of AGG, RB, and 

ASB in general at baseline, with fewer differences across sex observed in rates of change, 

consistent with prior work (e.g., Miller, Malone, & Dodge, 2010). However, we did not examine 

102 

 
potential interactions between sex and familial disadvantage, neighborhood disadvantage, 

parent-child conflict, or peer delinquency. Prior cross-sectional work has found associations 

between disadvantage and externalizing to persist regardless of whether sex was included as a 

covariate (Carroll et al., 2023), but it is unclear whether the same would be true longitudinally. 

There is some evidence that associations between the parent-child relationship and trajectories 

of ASB are similar for males and females (Janssen, Eichelsheim, Dekovic, & Bruinsma, 2017; 

Snyder, Schrepferman, Bullard, McEachern, & Patterson, 2012), but these studies followed 

participants across relatively brief windows of development. Lastly, regarding peer delinquency, 

some studies have found associations between peer influences and ASB to be stronger for 

males (e.g., Piquero, Gover, MacDonald, & Piquero, 2005), whereas others have found females 

to be more susceptible to the influences of delinquent peers (e.g., Haynie, Doogan, & Soller, 

2014). Haynie and colleagues also reported sex differences in the respective roles of selection 

and socialization, with stronger evidence of peer selection effects for girls. In short, much 

remains unknown about sex differences in the developmental origins of ASB. Future longitudinal 

studies should leverage variable-centered and person-centered methods, as well as quasi-

experimental designs, to examine interactions between sex and other predictors of AGG and 

RB trajectories. 

Lastly, none of the present studies examined the role of genotype-by-environment 

interaction (GxE) in the development of ASB. GxE refers to the differential impact of 

environmental risk or protective factors as a function of genotype (and vice versa, i.e., 

environmental factors impact the expression of genetic predispositions) (Plomin, DeFries, & 

Loehlin, 1977). Extant cross-sectional work has implicated GxE as an important contributor to 

youth ASB, particularly RB. For example, Burt and colleagues found shared environmental 

influences on RB to be amplified, and genetic contributions suppressed, for youth residing in 

disadvantaged neighborhoods (Burt, Klump, Gorman-Smith, & Neiderhiser, 2016). This pattern 

of findings, which has been consistent across studies (Cleveland, 2003; Tuvblad, Grann, & 

103 

 
Lichtenstein, 2006), is indicative of a bioecological GxE, with deleterious environments exerting 

such a strong effect on development that poor mental health outcomes are observed even 

among youth who are not genetically predisposed (Bronfenbrenner & Ceci, 1994).  

Despite these robust and consistent findings linking GxE to the occurrence of ASB, the 

role of GxE in the development of ASB is unknown, as models to assess GxE longitudinally do 

not yet exist (though they are in progress). Given age-related changes in both the frequency 

(Moffitt, 1993) and the etiology (Rhee & Waldman, 2002) of ASB, there are a variety of ways in 

which longitudinal findings may differ from cross-sectional ones. First, the manifestation of 

bioecological GxE may shift from one involving amplification of shared environmental influences 

to an increase in the effects of the nonshared environment. This developmental shift would be 

consistent with the age-related increase in nonshared environmental influences, and co-

occurring decrease in shared environmental effects, on ASB reported in prior meta-analytic 

work (Rhee & Waldman, 2002), which is hypothesized to reflect adolescents’ increasing ability 

to select their own environments outside the family. As youth progress into emerging adulthood, 

they may thus be less susceptible to familial influences but more affected by experiences in the 

neighborhood context that are unique to each person in a given family. Alternatively, GxE may 

follow an entirely different pattern. The diathesis-stress model of GxE posits that genetic risk 

would be activated in deleterious environments (Ingram & Luxton, 2005), meaning that genetic 

contributions to ASB would be expected to be amplified in disadvantaged neighborhoods. This 

pattern of GxE stands in direct contrast to that proposed by the bioecological model and 

suggests another way in which genetic and environmental influences may interact to shape 

youth behavioral trajectories. A shift from bioecological GxE during childhood to diathesis-stress 

GxE in adolescence would also help to explain age-related increases in the heritability of RB 

(Burt, 2015). In short, longitudinal studies are needed to understand GxE as a developmental 

process unfolding over time, rather than a static phenomenon. Future work should leverage 

both variable-centered and person-centered approaches to elucidate the effects of 

104 

 
neighborhood disadvantage on the etiology of the intercept and slope of ASB, as well as 

trajectory group membership. 

Despite these limitations, the present studies yield several important conclusions 

regarding the development of ASB. First, behavioral trajectories are shaped by genetic and 

nonshared environmental influences that are partially continuous across almost all of early 

development. That is, some of the same etiological factors that contribute to ASB in emerging 

adulthood are present as early as the preschool years. Some of these factors are person-

specific environmental, as the nonshared environment comprised not only transient, 

unsystematic effects but also enduring influences that shaped the magnitude of age-related 

change. Regarding specific predictors of ASB trajectories, the present studies implicated 

individual (i.e., biological sex), familial (i.e., parent-child conflict, household income), social 

network (i.e., peer delinquency), and neighborhood (i.e., concentrated deprivation) 

characteristics, consistent with theoretical work viewing development to unfold via interactions 

between the individual and numerous proximal and distal environmental contexts 

(Bronfenbrenner, 1988). 

Second, the nuanced ways in which Studies 2 and 3 were able to illuminate the 

predictors of desistance and escalation, respectively, underscore the need to leverage both 

variable-centered and person-centered statistical methods when modeling trajectories of 

psychopathology. Across all three studies, variable-centered methods were severely limited in 

their ability to model predictors of the slope, with neighborhood disadvantage and parent-child 

conflict found to predict declining ASB and household income found to predict an age-related 

increase. It was only by incorporating person-centered models that we were able to accurately 

conceptualize the respective roles of socioeconomic disadvantage and parent-child conflict in 

predicting rates of change. At the same time, taxometric work has consistently found ASB to 

exist as a continuous trait (Walters, 2011; Walters & Ruscio, 2013), like other forms of 

psychopathology (Krueger & Piasecki, 2002). Indeed, participants in the present studies were 

105 

 
not assigned to trajectory groups with absolute certainty, in part because the groups likely 

represent approximations of participants’ trajectories rather than qualitatively distinct entities. By 

leveraging both approaches together, the present studies were able to identify risk/protective 

factors associated with specific ASB trajectories, while simultaneously modeling the 

development of ASB in a way that was consistent with its underlying distribution. Future studies 

that make use of both statistical approaches, in conjunction with genetically informative designs 

that examine GxE, will further elucidate the origins of ASB and other forms of developmental 

psychopathology.  

106 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Tables 

Table 4.1. Cross-tabulation results for AGG and RB trajectory group membership 

RB 

Normative 
Escalating 
Desisting 
Persisting 
Total N 

Normative 
633 
100 
287 
58 
1078 

AGG 
Desisting 
168 
29 
322 
161 
680 

Persisting 
6 
20 
80 
184 
290 

Total N 

807 
149 
689 
403 
2048 

107 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
BIBLIOGRAPHY 

Achenbach, T. M., McConaughy, S. H., & Howell, C. T. (1987). Child/adolescent behavioral and 

emotional problems: implications of cross-informant correlations for situational 
specificity. Psychological bulletin, 101(2), 213. 

Achenbach, T. M., & Rescorla, L. A. (2001). Manual for the ASEBA school-age forms & 

profiles. Burlington, VT: University of Vermont Research Centre for Children, Youth and 
Families. 

Akaike, H. (1987). Factor analysis and AIC. In Selected Papers of Hirotugu Akaike (pp. 371- 

386). Springer, New York, NY. 

Alink, L. R., & Egeland, B. (2013). The roles of antisocial history and emerging adulthood 
developmental adaptation in predicting adult antisocial behavior. Aggressive 
behavior, 39(2), 131-140. 

American Psychiatric Association. (2013). Diagnostic and statistical manual of mental 
disorders (5th ed.). https://doi.org/10.1176/appi.books.9780890425596 

Arım, R. G., Dahinten, V. S., Marshall, S. K., & Shapka, J. D. (2011). An examination of the 
reciprocal relationships between adolescents’ aggressive behaviors and their 
perceptions of parental nurturance. Journal of Youth and Adolescence, 40, 207-220. 

Barker, E. D., Séguin, J. R., White, H. R., Bates, M. E., Lacourse, E., Carbonneau, R., & 

Tremblay, R. E. (2007). Developmental trajectories of male physical violence and theft: 
Relations to neurocognitive performance. Archives of General Psychiatry, 64(5), 592–
599.  

Bartels, M., Van den Oord, E. J. C. G., Hudziak, J. J., Rietveld, M. J. H., Van Beijsterveldt, C. E. 
M., & Boomsma, D. I. (2004). Genetic and environmental mechanisms underlying 
stability and change in problem behaviors at ages 3, 7, 10, and 12. Developmental 
psychology, 40(5), 852. 

Bauer, D. J., & Curran, P. J. (2004). The integration of continuous and discrete latent variable 

models: potential problems and promising opportunities. Psychological methods, 9(1), 3. 

Beam, C. R., & Turkheimer, E. (2013). Phenotype–environment correlations in longitudinal twin 

models. Development and psychopathology, 25(1), 7-16. 

Beauchaine, T. P., Webster-Stratton, C., & Reid, M. J. (2005). Mediators, moderators, and 

predictors of 1-year outcomes among children treated for early-onset conduct problems: 
a latent growth curve analysis. Journal of consulting and clinical psychology, 73(3), 371. 

Becht, A. I., Prinzie, P., Deković, M., van den Akker, A. L., & Shiner, R. L. (2016). Child 

personality facets and overreactive parenting as predictors of aggression and rule-
breaking trajectories from childhood to adolescence. Development and 
Psychopathology, 28(2), 399–413.  

Becker, A., Woerner, W., Hasselhorn, M., Banaschewski, T., & Rothenberger, A. (2004). 

108 

 
 
 
 
 
 
 
 
 
 
 
 
 
Validation of the parent and teacher SDQ in a clinical sample. European child & 
adolescent psychiatry, 13(2), ii11-ii16. 

Benson, M. J., & Buehler, C. (2012). Family process and peer deviance influences on 

adolescent aggression: Longitudinal effects across early and middle adolescence. Child 
development, 83(4), 1213-1228. 

Bentler, P. M., & Bonett, D. G. (1980). Significance tests and goodness of fit in the analysis of 

covariance structures. Psychological bulletin, 88(3), 588. 

Berkout, O. V., Young, J. N., & Gross, A. M. (2011). Mean girls and bad boys: Recent research 
on gender differences in conduct disorder. Aggression and Violent Behavior, 16(6), 503-
511. 

Bollen, K. A. (1989). A new incremental fit index for general structural equation 

models. Sociological methods & research, 17(3), 303-316. 

Bongers, I. L., Koot, H. M., Van Der Ende, J., & Verhulst, F. C. (2004). Developmental 

trajectories of externalizing behaviors in childhood and adolescence. Child 
development, 75(5), 1523-1537. 

Brame, B., Nagin, D. S., & Tremblay, R. E. (2001). Developmental trajectories of physical 
aggression from school entry to late adolescence. Journal of child psychology and 
psychiatry, 42(4), 503-512. 

Broidy, L. M., Nagin, D. S., Tremblay, R. E., Bates, J. E., Brame, B., Dodge, K. A., Fergusson, 

D., Horwood, J. L., Loeber, R., Laird, R., Lynam, D. R., Moffitt, T. E., Pettit, G. S., & 
Vitaro, F. (2003). Developmental trajectories of childhood disruptive behaviors and 
adolescent delinquency: A six-site, cross-national study. Developmental Psychology, 
39(2), 222–245.  

Bronfenbrenner, U. (1988). Interacting systems in human development. Research paradigms: 

Present and future. Persons in context: Developmental processes, 2, 25-49. 

Bronfenbrenner, U., & Ceci, S. J. (1994). Nature-nurture reconceptualized in developmental 

perspective: A bioecological model. Psychological review, 101(4), 568. 

Burt, S. A. (2009). Rethinking environmental contributions to child and adolescent 

psychopathology: a meta-analysis of shared environmental influences. Psychological 
bulletin, 135(4), 608. 

Burt, S. A. (2012). How do we optimally conceptualize the heterogeneity within antisocial 
behavior? An argument for aggressive versus non-aggressive behavioral 
dimensions. Clinical psychology review, 32(4), 263-279. 

Burt, S. A. (2013). Do etiological influences on aggression overlap with those on rule breaking? 

A meta-analysis. Psychological Medicine, 43(9), 1801-1812. 

Burt, S. A. (2015). Evidence that the gene–environment interactions underlying youth conduct 

problems vary across development. Child Development Perspectives, 9(4), 217-221. 

109 

 
 
 
 
 
 
 
 
 
 
 
 
 
Burt, S. A. (2022). The genetic, environmental, and cultural forces influencing youth antisocial 

behavior are tightly intertwined. Annual review of clinical psychology, 18, 155-178. 

Burt, S. A., Clark, D. A., Gershoff, E. T., Klump, K. L., & Hyde, L. W. (2021). Twin Differences in 
Harsh Parenting Predict Youth’s Antisocial Behavior. Psychological science, 32(3), 395-
409. 

Burt, S. A., Brent Donnellan, M., Slawinski, B. L., & Klump, K. L. (2016). The phenomenology of 
non-aggressive antisocial behavior during childhood. Journal of abnormal child 
psychology, 44, 651-661. 

Burt, S. A., Klahr, A. M., & Klump, K. L. (2015). Do non-shared environmental influences persist 
over time? An examination of days and minutes. Behavior genetics, 45(1), 24-34. 

Burt, S. A., & Klump, K. L. (2013). The Michigan state university twin registry (MSUTR): An 

update. Twin Research and Human Genetics, 16(1), 344-350. 

Burt, S. A., & Klump, K. L. (2019). The Michigan State University Twin Registry (MSUTR): 15 

years of twin and family research. Twin research and human genetics: the official journal 
of the International Society for Twin Studies, 22(6), 741. 

Burt, S. A., Klump, K. L., Gorman-Smith, D., & Neiderhiser, J. M. (2016). Neighborhood 

disadvantage alters the origins of children’s nonaggressive conduct problems. Clinical 
psychological science, 4(3), 511-526. 

Burt, S. A., Krueger, R. F., McGue, M., & Iacono, W. (2003). Parent-child conflict and the 
comorbidity among childhood externalizing disorders. Archives of general 
psychiatry, 60(5), 505-513. 

Burt, S. A., McGue, M., Carter, L. A., & Iacono, W. G. (2007). The different origins of stability 

and change in antisocial personality disorder symptoms. Psychological Medicine, 37(1), 
27. 

Burt, S. A., McGue, M., & Iacono, W. G. (2009). Nonshared environmental mediation of the 

association between deviant peer affiliation and adolescent externalizing behaviors over 
time: results from a cross-lagged monozygotic twin differences design. Developmental 
psychology, 45(6), 1752. 

Burt, S. A., McGue, M., Iacono, W. G., & Krueger, R. F. (2006). Differential parent-child 

relationships and adolescent externalizing symptoms: cross-lagged analyses within a 
monozygotic twin differences design. Developmental psychology, 42(6), 1289. 

Burt, S. A., Slawinski, B. L., Carsten, E. E., Harden, K. P., Hyde, L. W., & Klump, K. L. (2019). 
How should we understand the absence of sex differences in the genetic and 
environmental origins of antisocial behavior?. Psychological medicine, 49(10), 1600. 

Carroll, S. L., Mikhail, M. E., & Burt, S. A. (2023). The development of youth antisocial behavior 

across time and context: A systematic review and integration of person-centered and 
variable-centered research. Clinical psychology review, 102253. 

110 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
Carroll, S. L., Shewark, E. A., Mikhail, M. E., Thaler, D. J., Pearson, A. L., Klump, K. L., & Burt, 

S. A. (2023). Identifying the ‘active ingredients’ of socioeconomic disadvantage for youth 
outcomes in middle childhood. Development and Psychopathology, 1-9. 

Cleveland, H. H. (2003). Disadvantaged neighborhoods and adolescent aggression: Behavioral 
genetic evidence of contextual effects. Journal of Research on Adolescence, 13(2), 211-
238. 

Conger, R. D., Ge, X., Elder Jr, G. H., Lorenz, F. O., & Simons, R. L. (1994). Economic stress, 

coercive family process, and developmental problems of adolescents. Child 
development, 65(2), 541-561. 

Copeland, W. E., Miller-Johnson, S., Keeler, G., Angold, A., & Costello, E. J. (2007). Childhood 

psychiatric disorders and young adult crime: A prospective, population-based 
study. American Journal of psychiatry, 164(11), 1668-1675. 

Côté, S., Vaillancourt, T., LeBlanc, J. C., Nagin, D. S., & Tremblay, R. E. (2006). The 

development of physical aggression from toddlerhood to pre-adolescence: A nation wide 
longitudinal study of Canadian children. Journal of abnormal child psychology, 34, 68-82. 

DeGarmo, D. S., & Forgatch, M. S. (2005). Early development of delinquency within divorced 
families: Evaluating a randomized preventive intervention trial. Developmental 
science, 8(3), 229-239. 

DeGarmo, D. S., Patterson, G. R., & Forgatch, M. S. (2004). How do outcomes in a specified 

parent training intervention maintain or wane over time?. Prevention Science, 5, 73-89. 

de Haan, A. D., Prinzie, P., & Deković, M. (2012). Change and reciprocity in adolescent 
aggressive and rule-breaking behaviors and parental support and dysfunctional 
discipline. Development and Psychopathology, 24(1), 301–315.  

Diamantopoulou, S., Verhulst, F. C., & van der Ende, J. (2011). Gender differences in the 

development and adult outcome of co-occurring depression and delinquency in 
adolescence. Journal of Abnormal Psychology, 120(3), 644–655.  

Doornwaard, S. M., Branje, S., Meeus, W. H. J., & ter Bogt, T. F. M. (2012). Development of 

adolescents’ peer crowd identification in relation to changes in problem behaviors. 
Developmental Psychology, 48(5), 1366–1380.  

Dong, L. (2016). The role of perinatal depression in the development of co-occurring aggression 
and anxiety/depression problems [PhD Thesis, Emory University]. In Dissertation 
Abstracts International: Section B: The Sciences and Engineering.  

Dugré, J. R., Giguère, C. É., & Potvin, S. (2024). The company you keep: the neglected role of 

affiliating with delinquent friends in the development of the cannabis-violence 
link. Addictive behaviors, 151, 107939. 

Eamon, M. K. (2002). Poverty, parenting, peer, and neighborhood influences on young 
adolescent antisocial behavior. Journal of Social Service Research, 28(1), 1-23. 

111 

 
 
 
 
 
 
 
 
 
Ehrenreich, S. E., Beron, K. J., Brinkley, D. Y., & Underwood, M. K. (2014). Family predictors of 

continuity and change in social and physical aggression from ages 9 to 18. Aggressive 
Behavior, 40(5), 421–439.  

Eley, T. C., Lichtenstein, P., & Moffitt, T. E. (2003). A longitudinal behavioral genetic analysis of 

the etiology of aggressive and nonaggressive antisocial behavior. Development and 
psychopathology, 15(2), 383-402. 

Elkins, I. J., McGue, M., & Iacono, W. G. (1997). Genetic and environmental influences on 
parent–son relationships: Evidence for increasing genetic influence during 
adolescence. Developmental psychology, 33(2), 351. 

Embretson, S. E., & Reise, S. P. (2013). Item response theory. Psychology Press. 

Enders, C. K., & Bandalos, D. L. (2001). The relative performance of full information maximum 
likelihood estimation for missing data in structural equation models. Structural equation 
modeling, 8(3), 430-457. 

El-Sheikh, M., & Elmore–Staton, L. O. R. I. (2004). The link between marital conflict and child 

adjustment: Parent–child conflict and perceived attachments as mediators, potentiators, 
and mitigators of risk. Development and Psychopathology, 16(3), 631-648. 

Ettekal, I., & Ladd, G. W. (2015). Developmental pathways from childhood aggression–

disruptiveness, chronic peer rejection, and deviant friendships to early‐adolescent rule 
breaking. Child development, 86(2), 614-631. 

Ferdinand, R. F., Hoogerheide, K. N., Van Der Ende, J., Visser, J. H., Koot, H. M., Kasius, M. 

C., & Verhulst, F. C. (2003). The role of the clinician: three‐year predictive value of 
parents’, teachers’, and clinicians’ judgment of childhood psychopathology. Journal of 
Child Psychology and Psychiatry, 44(6), 867-876. 

Fontaine, N. M., Lacourse, E., Vitaro, F., & Tremblay, R. E. (2014). Links between trajectories of 

self‐reported violent and nonviolent offending and official offending during adolescence 
and adulthood. Criminal behaviour and mental health, 24(4), 277-290. 

Ganiban, J. M., Saudino, K. J., Ulbricht, J., Neiderhiser, J. M., & Reiss, D. (2008). Stability and 
change in temperament during adolescence. Journal of personality and social 
psychology, 95(1), 222. 

Gard, A. M., Waller, R., Shaw, D. S., Forbes, E. E., Hariri, A. R., & Hyde, L. W. (2017). The long 
reach of early adversity: Parenting, stress, and neural pathways to antisocial behavior in 
adulthood. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 2(7), 582-
590. 

Gardner, T. W. (2006). Developmental trajectories of antisocial behavior: Prediction of academic 

failure, peer rejection, and depression in adolescence. University of Oregon. 

Gardner, T. W., Dishion, T. J., & Connell, A. M. (2008). Adolescent self-regulation as resilience: 

Resistance to antisocial behavior within the deviant peer context. Journal of abnormal 
child psychology, 36(2), 273-284. 

112 

 
 
 
 
 
 
 
 
Goldsmith, H. H. (1993). Nature-nurture issues in the behavioral genetic context: Overcoming 

barriers to communication. Nature, nurture, and psychology, 325-339. 

Goodman, R. (2001). Strengths and Difficulties Questionnaire (SDQ). Youthinmind, London. 

Retreived 20 August 2007. 

Grimm, K. J., Ram, N., Estabrook, R. (2017). Growth modeling: Structural equation and 

multilevel modeling approaches. The Guilford Press. 

Hagen, K. A., Ogden, T., & Bjørnebekk, G. (2011). Treatment outcomes and mediators of 
parent management training: A one-year follow-up of children with conduct 
problems. Journal of Clinical Child & Adolescent Psychology, 40(2), 165-178. 

Hair, J. F., Anderson, R. E., Tatham, R. L., & Black, W. C. (1998). Factor analysis. Multivariate 

data analysis. NJ: Prentice-Hall, 3, 98-99. 

Haynie, D. L., Doogan, N. J., & Soller, B. (2014). Gender, friendship networks, and delinquency: 

A dynamic network approach. Criminology, 52(4), 688-722. 

He, J. P., Burstein, M., Schmitz, A., & Merikangas, K. R. (2013). The Strengths and Difficulties 

Questionnaire (SDQ): the factor structure and scale validation in US 
adolescents. Journal of abnormal child psychology, 41(4), 583-595. 

Hirschi, T., & Gottfredson, M. R. (1983). Age and the Explanation of Crime. American Journal of 

Sociology, 89, 552–584. 

Hooper, D., Coughlan, J., & Mullen, M. (2008). Evaluating model fit: a synthesis of the structural 
equation modelling literature. In 7th European Conference on research methodology for 
business and management studies (Vol. 2008, pp. 195-200). 

Hopwood, C. J., Donnellan, M. B., Blonigen, D. M., Krueger, R. F., McGue, M., Iacono, W. G., & 
Burt, S. A. (2011). Genetic and environmental influences on personality trait stability and 
growth during the transition to adulthood: a three-wave longitudinal study. Journal of 
personality and social psychology, 100(3), 545. 

Hox, J., & Stoel, R. D. (2014). Multilevel and SEM approaches to growth curve modeling. Wiley 

StatsRef: Statistics Reference Online. 

Hu, L. T., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: 
Conventional criteria versus new alternatives. Structural equation modeling: a 
multidisciplinary journal, 6(1), 1-55. 

Ingoldsby, E. M. (2003). Neighborhood contextual factors and early-starting antisocial behavior 
[PhD Thesis]. In Dissertation Abstracts International: Section B: The Sciences and 
Engineering.  

Ingram, R. E., & Luxton, D. D. (2005). Vulnerability-stress models. Development of 
psychopathology: A vulnerability-stress perspective, 46(2), 32-46. 

113 

 
 
 
 
 
 
 
 
 
Isen, J., Tuvblad, C., Younan, D., Ericson, M., Raine, A., & Baker, L. A. (2022). Developmental 

trajectories of delinquent and aggressive behavior: Evidence for differential 
heritability. Child Psychiatry & Human Development, 53(2), 199-211. 

Jacobson, K. C., Prescott, C. A., & Kendler, K. S. (2002). Sex differences in the genetic and 

environmental influences on the development of antisocial behavior. Development and 
Psychopathology, 14(2), 395-416. 

Janssen, H. J., Eichelsheim, V. I., Deković, M., & Bruinsma, G. J. (2017). Sex differences in 
longitudinal pathways from parenting to delinquency. European Journal on Criminal 
Policy and Research, 23, 503-521. 

Karriker-Jaffe, K. J., Foshee, V. A., Ennett, S. T., & Suchindran, C. (2008). The development of 

aggression during adolescence: Sex differences in trajectories of physical and social 
aggression among youth in rural areas. Journal of Abnormal Child Psychology, 36(8), 
1227–1236.  

Kendler, K. S., Jacobson, K., Myers, J. M., & Eaves, L. J. (2008). A genetically informative 

developmental study of the relationship between conduct disorder and peer deviance in 
males. Psychological medicine, 38(7), 1001-1011. 

Kessler, R. C., Chiu, W. T., Demler, O., & Walters, E. E. (2005). Prevalence, severity, and 

comorbidity of 12-month DSM-IV disorders in the National Comorbidity Survey 
Replication. Archives of general psychiatry, 62(6), 617-627. 

Kind, A. J., & Buckingham, W. R. (2018). Making neighborhood-disadvantage metrics 

accessible—the neighborhood atlas. The New England journal of medicine, 378(26), 
2456. 

Kjøbli, J., Hukkelberg, S., & Ogden, T. (2013). A randomized trial of group parent training: 

Reducing child conduct problems in real-world settings. Behaviour Research and 
Therapy, 51(3), 113– 121. 

Klahr, A. M., Klump, K. L., & Burt, S. A. (2014). The etiology of the association between child 

antisocial behavior and maternal negativity varies across aggressive and non-aggressive 
rule-breaking forms of antisocial behavior. Journal of abnormal child psychology, 42, 
1299-1311. 

Krueger, R. F., & Piasecki, T. M. (2002). Toward a dimensional and psychometrically-informed 
approach to conceptualizing psychopathology. Behaviour research and therapy, 40(5), 
485-499. 

Kupersmidt, J. B., Griesler, P. C., DeRosier, M. E., Patterson, C. J., & Davis, P. W. (1995). 
Childhood aggression and peer relations in the context of family and neighborhood 
factors. Child development, 66(2), 360-375. 

Lacourse, E., Boivin, M., Brendgen, M., Petitclerc, A., Girard, A., Vitaro, F., ... & Tremblay, R. E. 
(2014). A longitudinal twin study of physical aggression during early childhood: evidence 
for a developmentally dynamic genome. Psychological medicine, 44(12), 2617. 

114 

 
 
 
 
 
 
Lacourse, E., Dupéré, V., & Loeber, R. (2008). Developmental trajectories of violence and theft. 

In R. Loeber, D. P. Farrington, M. Stouthamer-Loeber, & H. R. White (Eds.), Violence 
and serious theft: Development and prediction from childhood to adulthood BT - 
Violence and serious theft: Development and prediction from childhood to adulthood (pp. 
231–268, Chapter xxiii, 404 Pages). Routledge/Taylor & Francis Group, New York, NY.  

Lahey, B. B., Loeber, R., Burke, J. D., & Applegate, B. (2005). Predicting future antisocial 

personality disorder in males from a clinical assessment in childhood. Journal of 
consulting and clinical psychology, 73(3), 389. 

Lee, Y., Liu, X., & Watson, M. W. (2016). The timing effect of bullying in childhood and 

adolescence on developmental trajectories of externalizing behaviors. Journal of 
interpersonal violence, 31(17), 2775-2800. 

Leventhal, T., & Brooks-Gunn, J. (2000). The neighborhoods they live in: the effects of 

neighborhood residence on child and adolescent outcomes. Psychological 
bulletin, 126(2), 309. 

Little, R. J., & Rubin, D. B. (2019). Statistical analysis with missing data (Vol. 793). John Wiley & 

Sons. 

MacCallum, R. C. (2009). Factor Analysis. In R. E. Millsap, & A. Maydeu-Olivares (Eds.), The 
SAGE Handbook of Quantitative Methods in Psychology (pp. 123-147). London: SAGE 
Publications Ltd. 

Maldonado-Molina, M. M., Reingle, J. M., Tobler, A. L., Jennings, W. G., & Komro, K. A. (2010). 
Trajectories of physical aggression among Hispanic urban adolescents and young 
adults: An application of latent trajectory modeling from ages 12 to 18. American Journal 
of Criminal Justice, 35, 121-133. 

Mann, F. D., Kretsch, N., Tackett, J. L., Harden, K. P., & Tucker-Drob, E. M. (2015). Person× 

environment interactions on adolescent delinquency: Sensation seeking, peer deviance 
and parental monitoring. Personality and individual differences, 76, 129-134. 

Marsh, H. W., & Hau, K. T. (2002). Multilevel modeling of longitudinal growth and change: 

Substantive effects or regression toward the mean artifacts?. Multivariate Behavioral 
Research, 37(2), 245-282. 

Martino, S. C., Ellickson, P. L., Klein, D. J., McCaffrey, D., & Edelen, M. O. (2008). Multiple 

trajectories of physical aggression among adolescent boys and girls. Aggressive 
behavior, 34(1), 61-75. 

Mata, A. D. (2013). Structured activity involvement and behavior problems: Investigation of 

selection and transactional effects with the use of variable and person-centered 
approaches [PhD Thesis]. In Dissertation Abstracts International: Section B: The 
Sciences and Engineering.  

Mata, A. D., & Van Dulmen, M. H. (2012). Adult-onset antisocial behavior trajectories: 
Associations with adolescent family processes and emerging adulthood 
functioning. Journal of interpersonal violence, 27(1), 177-193. 

115 

 
 
 
 
 
 
Maughan, B., Pickles, A., Rowe, R., Costello, E. J., & Angold, A. (2000). Developmental 

trajectories of aggressive and non-aggressive conduct problems. Journal of Quantitative 
Criminology, 16(2), 199–221.  

McArdle, J. J., & Epstein, D. (1987). Latent growth curves within developmental structural 

equation models. Child Development, 110–133. 

McGue, M., & Bouchard, T. J. (1984). Adjustment of twin data for the effects of age and 

sex. Behavior genetics, 14(4), 325-343. 

Miller, S., Malone, P. S., Dodge, K. A., & Conduct Problems Prevention Research Group. 

(2010). Developmental trajectories of boys’ and girls’ delinquency: Sex differences and 
links to later adolescent outcomes. Journal of abnormal child psychology, 38, 1021-
1032. 

Moffitt, T. E. (1993). Adolescence-Limited and Life-Course-Persistent Antisocial Behavior: A 

Developmental Taxonomy. 100(4), 674–701. 

Moffitt, T. E. (2003). Life-course-persistent and adolescence-limited antisocial behavior: A 10-

year research review and a research agenda. 

Moffitt, T. E., Caspi, A., Harrington, H., & Milne, B. J. (2002). Males on the life-course-persistent 

and adolescence-limited antisocial pathways: Follow-up at age 26 years. Development 
and psychopathology, 14(1), 179-207. 

Monahan, K. C., Steinberg, L., & Cauffman, E. (2009). Affiliation with antisocial peers, 
susceptibility to peer influence, and antisocial behavior during the transition to 
adulthood. Developmental psychology, 45(6), 1520. 

Mrug, S., Elliott, M., Gilliland, M. J., Grunbaum, J. A., Tortolero, S. R., Cuccaro, P., & Schuster, 
M. (2008). Positive parenting and early puberty in girls: protective effects against 
aggressive behavior. Archives of pediatrics & adolescent medicine, 162(8), 781-786. 

Muris, P., Meesters, C., & van den Berg, F. (2003). The strengths and difficulties questionnaire 

(SDQ). European child & adolescent psychiatry, 12(1), 1-8. 

Murray, A. L., Eisner, M., & Ribeaud, D. (2020). Within‐person analysis of developmental 

cascades between externalising and internalising problems. Journal of Child Psychology 
and Psychiatry, 61(6), 681–688.  

Muthén, B. (2004). Latent variable analysis. The Sage handbook of quantitative methodology for 

the social sciences, 345(368), 106-109. 

Muthén, B., & Asparouhov, T. (2002). Latent variable analysis with categorical outcomes: 
Multiple-group and growth modeling in Mplus. Mplus web notes, 4(5), 1-22. 

Muthén, B.O., du Toit, S.H.C., & Spisic, D. (1997). Robust inference using weighted-least 

squares and quadratic estimating equations in latent variable modeling with categorical 
and continuous outcomes. Psychometrika. 

116 

 
 
 
 
 
 
 
 
 
Muthén, B., & Muthén, L. K. (2000). Integrating person‐centered and variable‐centered 

analyses: Growth mixture modeling with latent trajectory classes. Alcoholism: Clinical 
and experimental research, 24(6), 882-891. 

Muthén, L. K., and B. O. Muthén. 1998-2019. Mplus User’s Guide. Eighth Edition. Los Angeles, 

CA: Muthén & Muthén. 

Nagin, D. S. (1999). Analyzing Developmental Trajectories: A Semiparametric, Group-Based 

Approach. Psychological Methods, 4(2), 139–157. 

Nagin, D. (2005). Group-based modeling of development. Harvard University Press. 

Nagin, D. S., Barker, T., Lacourse, E., & Tremblay, R. E. (2009). The inter-relationship of 

temporally distinct risk markers and the transition from childhood physical aggression to 
adolescent violent delinquency. Monatsschrift für Kriminologie und 
Strafrechtsreform, 92(2-3), 102-116. 

Nagin, D. S., & Tremblay, R. E. (2005). Developmental trajectory groups: Fact or a useful 

statistical fiction?. Criminology, 43(4), 873-904. 

Narusyte, J., Neiderhiser, J. M., Andershed, A. K., D'Onofrio, B. M., Reiss, D., Spotts, E., ... & 
Lichtenstein, P. (2011). Parental criticism and externalizing behavior problems in 
adolescents: The role of environment and genotype–environment correlation. Journal of 
abnormal psychology, 120(2), 365. 

Neale, M. C., & Cardon, L. R. (1992). NATO ASI series D: Behavioural and social sciences, Vol. 

67. Methodology for genetic studies of twins and families. 

NICHD Early Child Care Research Network, & Arsenio, W. F. (2004). Trajectories of physical 

aggression from toddlerhood to middle childhood: Predictors, correlates, and outcomes. 
Monographs of the Society for Research in Child Development, i–143. 

Odgers, C. L., Moffitt, T. E., Broadbent, J. M., Dickson, N., Hancox, R. J., Harrington, H., ... & 

Caspi, A. (2008). Female and male antisocial trajectories: From childhood origins to 
adult outcomes. Development and psychopathology, 20(2), 673. 

Olson, S. L., Sameroff, A. J., Lansford, J. E., Sexton, H., Davis-Kean, P., Bates, J. E., ... & 
Dodge, K. A. (2013). Deconstructing the externalizing spectrum: Growth patterns of 
overt aggression, covert aggression, oppositional behavior, impulsivity/inattention, and 
emotion dysregulation between school entry and early adolescence. Development and 
psychopathology, 25(3), 817-842. 

Padilla-Walker, L. M., Memmott-Elison, M. K., & Coyne, S. M. (2018). Associations between 

Prosocial and Problem Behavior from Early to Late Adolescence. Journal of Youth and 
Adolescence, 47(5), 961–975.  

Peeters, H., Van Gestel, S., Vlietinck, R., Derom, C., & Derom, R. (1998). Validation of a 

telephone zygosity questionnaire in twins of known zygosity. Behavior genetics, 28(3), 
159-163. 

117 

 
 
 
 
 
Peterson, R. D., & Krivo, L. J. (2009). Segregated spatial locations, race-ethnic composition, 

and neighborhood violent crime. The Annals of the American Academy of Political and 
Social Science, 623(1), 93-107. 

Pingault, J. B., Rijsdijk, F., Zheng, Y., Plomin, R., & Viding, E. (2015). Developmentally dynamic 
genome: Evidence of genetic influences on increases and decreases in conduct 
problems from early childhood to adolescence. Scientific reports, 5(1), 1-9. 

Piquero, N. L., Gover, A. R., MacDonald, J. M., & Piquero, A. R. (2005). The influence of 

delinquent peers on delinquency: Does gender matter?. Youth & Society, 36(3), 251-
275. 

Plomin, R., & Daniels, D. (1987). Why are children in the same family so different from one 

another?. Behavioral and brain Sciences, 10(1), 1-16. 

Plomin, R., DeFries, J. C., & Loehlin, J. C. (1977). Genotype-environment interaction and 

correlation in the analysis of human behavior. Psychological bulletin, 84(2), 309. 

Porsch, R. M., Middeldorp, C. M., Cherny, S. S., Krapohl, E., Van Beijsterveldt, C. E., Loukola, 

A., ... & Bartels, M. (2016). Longitudinal heritability of childhood aggression. American 
Journal of Medical Genetics Part B: Neuropsychiatric Genetics, 171(5), 697-707. 

Preacher, K. J., Wichman, A. L., MacCallum, R. C., & Briggs, N. E. (2008). Latent growth curve 

modeling. Thousand Oaks, CA: Sage Publications. 

Putnick, D. L., & Bornstein, M. H. (2016). Measurement invariance conventions and reporting: 

The state of the art and future directions for psychological research. Developmental 
review, 41, 71-90. 

Raftery, A. E. (1995). Bayesian model selection in social research. Sociological methodology, 

111-163. 

Rescorla, L., Achenbach, T., Ivanova, M. Y., Dumenci, L., Almqvist, F., Bilenberg, N., ... & 

Verhulst, F. (2007). Behavioral and emotional problems reported by parents of children 
ages 6 to 16 in 31 societies. Journal of Emotional and behavioral Disorders, 15(3), 130-
142. 

Rhee, S. H., & Waldman, I. D. (2002). Genetic and environmental influences on antisocial 

behavior: a meta-analysis of twin and adoption studies. Psychological bulletin, 128(3), 
490. 

Rhemtulla, M., & Hancock, G. R. (2016). Planned missing data designs in educational 

psychology research. Educational Psychologist, 51(3-4), 305-316. 

Sacco, P., Bright, C. L., Jun, H.-J., & Stapleton, L. M. (2015). Developmental relations between 

alcohol and aggressive behavior among adolescents: Neighborhood and 
sociodemographic correlates. International Journal of Mental Health and Addiction, 
13(5), 603–617.  

Samek, D. R., Goodman, R. J., Erath, S. A., McGue, M., & Iacono, W. G. (2016). Antisocial peer 
affiliation and externalizing disorders in the transition from adolescence to young 

118 

 
 
 
 
 
 
 
 
 
 
 
 
adulthood: Selection versus socialization effects. Developmental psychology, 52(5), 813. 

Schaeffer, C. M., Petras, H., Ialongo, N., Masyn, K. E., Hubbard, S., Poduska, J., & Kellam, S. 

(2006). A comparison of girls' and boys' aggressive-disruptive behavior trajectories 
across elementary school: prediction to young adult antisocial outcomes. Journal of 
consulting and clinical psychology, 74(3), 500. 

Schaeffer, C. M., Petras, H., Ialongo, N., Poduska, J., & Kellam, S. (2003). Modeling growth in 
boys' aggressive behavior across elementary school: links to later criminal involvement, 
conduct disorder, and antisocial personality disorder. Developmental psychology, 39(6), 
1020. 

Schwartz, J. A., Solomon, S. J., & Valgardson, B. A. (2019). Socialization, selection, or both? 
The role of gene–environment interplay in the association between exposure to 
antisocial peers and delinquency. Journal of quantitative criminology, 35, 1-26. 

Sclove, S. L. (1987). Application of model-selection criteria to some problems in multivariate 

analysis. Psychometrika, 52(3), 333-343. 

Singh, G. K. (2003). Area deprivation and widening inequalities in US mortality, 1969–

1998. American journal of public health, 93(7), 1137-1143 

Snyder, J. J., Schrepferman, L. P., Bullard, L., McEachern, A. D., & Patterson, G. R. (2012). 

Covert antisocial behavior, peer deviancy training, parenting processes, and sex 
differences in the development of antisocial behavior during childhood. Development 
and psychopathology, 24(3), 1117-1138. 

Snyder, H. R., Young, J. F., & Hankin, B. L. (2019). Chronic stress exposure and generation are 

related to the p-factor and externalizing specific psychopathology in youth. Journal of 
Clinical Child & Adolescent Psychology, 48(2), 306-315. 

Spano, R., Rivera, C., & Bolland, J. M. (2010). Are chronic exposure to violence and chronic 

violent behavior closely related developmental processes during adolescence? Criminal 
Justice and Behavior, 37(10), 1160–1179.  

Steiger, J. H. (2007). Understanding the limitations of global fit assessment in structural 
equation modeling. Personality and Individual differences, 42(5), 893-898. 

Sterba, S. K. (2014). Fitting nonlinear latent growth curve models with individually varying time 
points. Structural Equation Modeling: A Multidisciplinary Journal, 21(4), 630-647. 

Thaweekoon, T. (2006). Effects of exposure to community violence on adolescent adjustment 

problems [PhD Thesis]. In Dissertation Abstracts International: Section B: The Sciences 
and Engineering.  

Tung, I., & Lee, S. S. (2017). Latent trajectories of adolescent antisocial behavior: Serotonin 
transporter linked polymorphic region (5-HTTLPR) genotype influences sensitivity to 
perceived parental support. Development and Psychopathology, 29(1), 185-201. 

Turkheimer, E. (2000). Three laws of behavior genetics and what they mean. Current directions 

in psychological science, 9(5), 160-164. 

119 

 
 
 
 
 
 
 
Turkheimer, E., & Gottesman, I. I. (1991). Individual differences and the canalization of human 

behavior. Developmental Psychology, 27, 18-22. 

Tuvblad, C., Grann, M., & Lichtenstein, P. (2006). Heritability for adolescent antisocial behavior 
differs with socioeconomic status: gene–environment interaction. Journal of Child 
Psychology and Psychiatry, 47(7), 734-743. 

Tuvblad, C., Narusyte, J., Grann, M., Sarnecki, J., & Lichtenstein, P. (2011). The genetic and 

environmental etiology of antisocial behavior from childhood to emerging 
adulthood. Behavior genetics, 41(5), 629-640. 

van Beijsterveldt, C. E. M., Bartels, M., Hudziak, J. J., & Boomsma, D. I. (2003). Causes of 

stability of aggression from early childhood to adolescence: A longitudinal genetic 
analysis in Dutch twins. Behavior Genetics, 33(5), 591-605. 

van Lier, P. A. C., Vitaro, F., Barker, E. D., Koot, H. M., & Tremblay, R. E. (2009). 

Developmental links between trajectories of physical violence, vandalism, theft, and 
alcohol-drug use from childhood to adolescence. Journal of Abnormal Child Psychology, 
37(4), 481–492.  

Walden, B., McGue, M., Burt, S. A., & Elkins, I. (2004). Identifying shared environmental 

contributions to early substance use: the respective roles of peers and parents. Journal 
of abnormal psychology, 113(3), 440. 

Walters, G. D. (2011). The latent structure of life-course-persistent antisocial behavior: Is 

Moffitt's developmental taxonomy a true taxonomy?. Journal of Consulting and Clinical 
Psychology, 79(1), 96. 

Walters, G. D., & Ruscio, J. (2013). Trajectories of youthful antisocial behavior: Categories or 

continua? Journal of Abnormal Child Psychology, 41(4), 653–666.  

Wang, M. T., & Dishion, T. J. (2012). The trajectories of adolescents’ perceptions of school 

climate, deviant peer affiliation, and behavioral problems during the middle school 
years. Journal of research on adolescence, 22(1), 40-53. 

Widaman, K. F., Ferrer, E., & Conger, R. D. (2010). Factorial invariance within longitudinal 
structural equation models: Measuring the same construct across time. Child 
development perspectives, 4(1), 10-18. 

Windle, M. (2000). A latent growth curve model of delinquent activity among 
adolescents. Applied Developmental Science, 4(4), 193-207. 

Xu, Y., Zhou, Y., Zhao, J., Xuan, Z., Li, W., Han, L., & Liu, H. (2021). The relationship between 
shyness and aggression in late childhood: The multiple mediation effects of parent-child 
conflict and self-control. Personality and individual differences, 182, 111058. 

Yan, J., Schoppe‐Sullivan, S. J., & Beauchaine, T. P. (2021). Paternal antisociality and growth 

in child delinquent behaviors: Moderating effects of child sex and respiratory sinus 
arrhythmia. Developmental psychobiology, 63(5), 1466-1481. 

120 

 
 
 
 
 
 
 
 
Yoon, D., Yoon, S., Yoon, M., & Snyder, S. M. (2019). Developmental trajectories of deviant 

peer affiliation in adolescence: Associations with types of child maltreatment and 
substance use. Children and Youth Services Review, 105, 104446. 

Yuan, K. H., & Chan, W. (2016). Measurement invariance via multigroup SEM: Issues and 

solutions with chi-square-difference tests. Psychological methods, 21(3), 405. 

121 

 
 
 
APPENDIX A: SUPPLEMENTAL MATERIAL FOR CHAPTER 1 

Tables 

Table S1.1. Standardized variance estimates from univariate ACE model for CBCL Conduct 

Problems scale 

Variance estimates 
c 

a 

e 

Univariate 
Intercept 
Slope 
Note: Bold font and asterisk indicate that the estimate was significantly different than zero at 

.20* 
  .11 

.31* 
.35* 

.50* 
.54* 

p<.05. 

122 

 
 
 
 
 
 
 
 
Figures 

Figure S1.1. Path diagram of a bivariate twin model for CBCL Conduct Problems scale. The 

variance in the intercept and the slope is partitioned into additive genetic effects (A1 and A2), 

shared environmental effects (C1 and C2), and nonshared environmental effects (E1 and E2). 

For ease of presentation, this path diagram represents one twin in a pair. Standardized path 

estimates are squared to represent the proportion of variance accounted for. 

123 

 
 
 
 
APPENDIX B: SUPPLEMENTAL MATERIAL FOR CHAPTER 2 

Supplementary Methods 

Measures 

AGG 

Teacher reports of AGG (20 items) were obtained at the first two assessment waves via 

the Achenbach Teacher Report Form (TRF; Achenbach & Rescorla, 2001). Teachers rated the 

twins’ behaviors during the preceding six months using a three-point scale (0=never to 

2=often/mostly true). The teachers of 115 participants were not available for assessment 

(because the twins were home-schooled or because parental consents to contact the teachers 

were completed incorrectly, etc.). Our teacher participation rate across both subsamples was 

86% at Wave 1 and 60% at Wave 2, with teacher reports available for 1,551 and 453 

participants, respectively. In addition, the twins completed the Youth Self-Report (YSR) at 

Waves 2 and 3, reporting on their own AGG (17 items) during the preceding six months using 

the three-point scale described above (Achenbach & Rescorla, 2001). Reports were available 

from 99% and 83% of the twins at Waves 2 and 3, respectively. 

Data Analytic Strategy 

Item Response Theory (IRT) Analyses 

Analyses were conducted in Mplus 8.4 (Muthén & Muthén, 1998-2019) using the 

weighted least squares mean- and variance-adjusted (WLSMV) estimator. As the twins were 

nested within families, analyses accounted for the nonindependence of observations using the 

CLUSTER command. We made use of an IRT model, which relates an individual’s trait level (θ) 

to their performance on a series of items while accounting for item characteristics. Given the 

relatively low frequency of AGG in the sample, and consistent with prior IRT analyses of 

antisocial behavior in the TBED-C (Burt, Donnellan, Slawinski, & Klump, 2016), items were 

coded dichotomously, with responses of 1 or 2 on the ASEBA instruments collapsed into a 

single “present” category (i.e., 0=behavior absent, 1=behavior present). We then applied the 

124 

 
two-parameter logistic (2PL) model, which estimates two item characteristics: discrimination (α) 

and difficulty (β) (Embretson & Reise, 2013). Item discrimination indicates the extent to which 

endorsement of a given item relates to one’s trait level; the higher the discrimination, the more 

accurately the item assesses standing on the trait of interest. Item difficulty indicates the trait 

level needed to have a 50% chance of endorsing the item. A higher difficulty thus indicates that 

one would need a relatively high standing on the latent trait continuum to receive a 1. Such an 

item would be diagnostic of high levels of AGG. 

Initial models included all AGG items from each available informant (i.e., 18 items from 

the CBCL and 20 items from the TRF at Wave 1; 18 items from the CBCL, 20 from the TRF, 

and 17 from the YSR at Wave 2; 18 from the CBCL and 17 from the YSR at Wave 3). Each 

wave comprised one latent factor. Items were omitted from subsequent models based on their 

characteristics (i.e., low discrimination and/or low endorsement rate) as well as theoretical 

considerations (i.e., whether they assessed emotional dysregulation or verbal AGG rather than 

physical AGG). Model fit was evaluated with three indices: root mean square error of 

approximation (RMSEA), Comparative Fit Index (CFI), and Tucker-Lewis Index (TLI). RMSEA 

values below .06 and CFI and TLI values above .95 are considered to indicate good fit (Hu & 

Bentler, 1999). For each model, we examined modification indices to identify sources of misfit 

and added the suggested residual covariances between items to the model. AGG mean scores 

were computed based on participants’ responses to the items retained in the final IRT model.  

IRT Modeling 

Supplementary Results 

The initial IRT model at each wave included all ASEBA AGG items from all available 

informants. The second model omitted items assessing emotion dysregulation (e.g., sudden 

changes in mood) and/or verbal AGG (e.g., argues). The final model excluded these items as 

well as those with standardized discrimination values < .3 (Hair, Anderson, Tatham, & Black, 

1998; i.e., suspicious) or endorsement rates near zero (i.e., threatens others). Fit indices are 

125 

 
reported in Table S2.2, and parameter estimates from the first two models are reported in 

Tables S2.3 and S2.4, respectively. The final model at each wave retained eight items from 

each available informant. Fit was good at Waves 1 and 3, according to all fit indices. At Wave 2, 

the CFI and TLI indicated some degree of misfit, even after residual covariances were added, 

but the RMSEA was well below recommended cutoffs (Hooper, Coughlan, & Mullen, 2008; Hu & 

Bentler, 1999; Steiger, 2007), indicating good fit. Discrimination and difficulty parameter 

estimates from the final model are reported in Table S2.5. As expected, items assessing 

behaviors that could be considered illegal (e.g., destroys things belonging to others, physically 

attacks people) were more difficult to endorse than were those assessing less extreme 

behaviors (e.g., disobedient at home, hot temper). Teacher-reported items generally 

discriminated better than did other informants’ and were more difficult to endorse at Wave 1 

than parent-reported items, consistent with prior work identifying somewhat higher rates of 

psychopathology in parent reports relative to teacher reports (Ferdinand et al., 2003). Mean 

scores were computed based on the items retained in the final model. Scores were multiplied by 

24 to correspond to the highest number of items included at any wave (i.e., Wave 2) and then 

log-transformed to account for positive skew. 

126 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S2.1. Youth AGG by age 

Tables 

Age 
6 
7 
8 
9 
10 
11 
12 
13 
14 
15 
16 
17 
18 
19 
20-21 

Total N 
598 
410 
346 
342 
352 
119 
82 
92 
185 
234 
164 
136 
68 
36 
7 

Mean AGG (SD) 
1.84 (2.06) 
1.53 (1.71) 
1.74 (1.98) 
1.22 (1.61) 
1.34 (1.78) 
1.01 (1.38) 
.73 (1.22) 
.73 (1.29) 
.77 (1.42) 
.59 (1.12) 
.44 (.99) 
.69 (1.13) 
.53 (1.09) 
.28 (.70) 
.00 (.00) 

Range AGG 
0-8 
0-7 
0-8 
0-8 
0-8 
0-7 
0-7 
0-5 
0-8 
0-6 
0-6 
0-6 
0-5 
0-3 
0-0 

Note: AGG sum scores were computed based on maternal reports on eight CBCL items. 

127 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S2.2. IRT model fit 

Wave 1 

Wave 2 

Wave 3 

Model 1 
Model 2 
Model 3 

Model 1 
Model 2 
Model 3 

Model 1 
Model 2 
Model 3 

RMSEA (90% CI) 

.010 (.007-.013) 
.032 (.028-.036) 
.013 (.004-.019) 

.029 (.027-.031) 
.029 (.025-.033) 
.032 (.027-.036) 

.034 (.029-.040) 
.036 (.026-.045) 
.036 (.024-.048) 

CFI 

.996 
.977 
.997 

.841 
.901 
.905 

.912 
.947 
.960 

TLI 

.994 
.961 
.996 

.831 
.890 
.890 

.906 
.940 
.952 

Note: Model 1 contained all ASEBA AGG items from each available informant at a given wave. 

Model 2 contained all items assessing physical AGG. Model 3 was the final model, containing 

items that assessed physical AGG and had good psychometric properties. 

128 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S2.3. Standardized IRT parameter estimates from models containing all ASEBA AGG 

items 

Argues 
Defiant, talks 
back 
Cruelty, 
bullying, or 
meanness 
Demands 
attention 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 
Physically 
attacks 
people 
Screams 
Explosive 
and 
unpredictable 
Easily 
frustrated 
Stubborn, 
sullen, or 
irritable 
Sudden 
changes in 
mood 
Sulks 
Suspicious 
Teases 
Hot temper 
Threatens 
people 
Unusually 
loud 

α 
.22 
- 

.34 

.16 

.47 

.45 

.36 

.69 

.28 
.31 

.29 
- 

- 

.24 

.30 

.18 
.39 
.19 
.25 
.32 

.22 

Wave 1 

CBCL 

TRF 

YSR 

β 
-.28 
- 

1.06 

.41 

.96 

.90 

.00 

.96 

1.41 
1.04 

.82 
- 

- 

.20 

.61 

1.03 
1.62 
.91 
.27 
1.49 

.72 

Α 
.91 
.84 

.82 

.71 

.82 

.84 

- 

.91 

.81 
.86 

.80 
.87 

.76 

.87 

β 
.61 
1.11 

1.14 

.70 

1.52 

1.72 

- 

1.00 

1.60 
1.61 

1.86 
1.51 

1.09 

1.02 

.74 

1.20 

1.10 
1.86 
1.42 
1.41 
1.87 

1.29 

.72 
.81 
.72 
.87 
.88 

.70 

129 

α 
- 
- 

- 

- 

- 

- 

- 

- 

- 
- 

- 
- 

- 

- 

- 

- 
- 
- 
- 
- 

- 

β 
- 
- 

- 

- 

- 

- 

- 

- 

- 
- 

- 
- 

- 

- 

- 

- 
- 
- 
- 
- 

- 

 
 
 
 
Table S2.3 (cont’d) 

Wave 2 

CBCL 

TRF 

YSR 

Argues 
Defiant, talks 
back 
Cruelty, 
bullying, or 
meanness 
Demands 
attention 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 
Physically 
attacks 
people 
Screams 
Explosive 
and 
unpredictable 
Easily 
frustrated 
Stubborn, 
sullen, or 
irritable 
Sudden 
changes in 
mood 
Sulks 
Suspicious 
Teases 
Hot temper 
Threatens 
people 
Unusually 
loud 

α 
.43 
- 

.61 

.50 

.59 

.69 

.55 

.71 

.66 
.64 

.54 
- 

- 

.39 

.42 

.44 
.44 
.45 
.47 
.74 

.53 

β 
-.32 
- 

.92 

.38 

.91 

.67 

.00 

.61 

1.12 
1.03 

.70 
- 

- 

.01 

.22 

.60 
1.06 
.50 
.27 
1.07 

.42 

Α 
.75 
.93 

.89 

.75 

.71 

.78 

- 

.89 

.87 
.85 

.84 
.96 

.83 

.72 

.81 

.72 
.86 
.76 
.92 
1.00 

.83 

130 

β 
-.07 
.02 

.18 

.21 

.94 

.95 

- 

.19 

.67 
.87 

1.07 
.46 

.33 

.16 

.26 

.46 
.72 
.39 
.53 
.69 

.49 

α 
.25 
- 

.44 

.21 

.53 

.62 

.35 

.57 

.52 
.54 

.52 
- 

β 
-.42 
- 

.35 

.60 

.83 

1.18 

.19 

.70 

1.08 
1.18 

.62 
- 

- 

- 

.001 

-.11 

.28 

-.20 

- 
.30 
.20 
.44 
.52 

.36 

- 
.36 
.59 
-.01 
1.08 

-.04 

 
 
 
 
 
 
 
Table S2.3 (cont’d) 

Wave 3 

CBCL 

TRF 

YSR 

β 
- 
- 

α 
.41 
- 

β 
-.41 
- 

α 
.79 
- 

β 
-.42 
- 

Α 
- 
- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

.49 

.71 

.52 

.87 

.00 

.23 

.94 

.32 

.64 

.55 

.67 

.86 

.63 

.67 

.40 

.64 

.85 

.57 

1.58 

1.39 

1.05 

1.20 

Argues 
Defiant, talks 
back 
Cruelty, 
bullying, or 
meanness 
Demands 
attention 
Destroys own 
things 
Destroys 
things 
belonging to 
others 
Disobedient at 
home* 
Disobedient at 
school 
Many fights 
Physically 
attacks people 
Screams 
Explosive and 
unpredictable 
Easily 
frustrated 
Stubborn, 
sullen, or 
irritable 
Sudden 
changes in 
mood 
Sulks 
Suspicious 
Teases 
Hot temper 
Threatens 
people 
Unusually loud 
Note: *The CBCL item “disobedient at home” was used as the anchor item at each wave for 

.72 
1.13 
.66 
.21 
1.26 

- 
.53 
.54 
.07 
1.72 

- 
.32 
.31 
.59 
.35 

.65 
.45 
.46 
.85 
.86 

1.13 
- 

1.74 
1.15 

1.55 
1.71 

.71 
- 

.28 
- 

.46 
.70 

.70 
.86 

.95 
- 

- 
- 
- 
- 
- 

- 
- 
- 
- 
- 

1.10 

-.41 

-.13 

-.10 

.34 

.44 

.53 

.77 

.71 

.38 

.47 

.04 

.97 

.24 

- 
- 

- 
- 

- 
- 

- 
- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

- 

model identification purposes. 

131 

 
 
 
 
 
Table S2.4. Standardized IRT parameter estimates from models containing all ASEBA items 

assessing physical AGG 

CBCL 

TRF 

YSR 

Wave 1 

Cruelty, 
bullying, or 
meanness 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 
Physically 
attacks 
people 
Explosive 
and 
unpredictable 
Suspicious 
Hot temper 
Threatens 
people 

Cruelty, 
bullying, or 
meanness 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 

α 
.59 

.78 

.88 

.75 

.69 

.65 
.64 

- 

.61 
.58 
.64 

α 
.58 

.62 

.67 

.52 

.74 

.71 

β 
1.08 

.99 

.91 

.00 

1.05 

1.40 
1.05 

Α 
.88 

.70 

.75 

- 

.96 

.82 
.87 

β 
1.24 

1.64 

1.84 

- 

1.12 

1.71 
1.72 

- 

.83 

1.63 

1.65 
.26 
1.49 

.73 
.76 
.85 

1.98 
1.54 
1.99 

α 
- 

- 

- 

- 

- 

- 
- 

- 

- 
- 
- 

CBCL 

TRF 

YSR 

Wave 2 

β 
.18 

.94 

.91 

- 

.14 

.55 

α 
.50 

.53 

.61 

.30 

.64 

.58 

Β 
.91 

.81 

.63 

.00 

.50 

1.00 

Α 
.83 

.66 

.76 

- 

.87 

.90 

132 

β 
- 

- 

- 

- 

- 

- 
- 

- 

- 
- 
- 

β 
.23 

.77 

1.14 

.23 

.55 

.95 

 
 
 
 
 
 
 
Table S2.4 (cont’d) 

Physically 
attacks 
people 
Explosive 
and 
unpredictable 
Suspicious 
Hot temper 
Threatens 
people 

Cruelty, 
bullying, or 
meanness 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 
Physically 
attacks 
people 
Explosive 
and 
unpredictable 
Suspicious 
Hot temper 
Threatens 
people 

.67 

.94 

- 

.37 
.51 
.75 

α 
.78 

.71 

.70 

.83 

.66 

.80 
.91 

- 

.20 
.78 
.94 

CBCL 

- 

1.12 
.18 
1.00 

β 
.92 

1.14 

.99 

.00 

.91 

1.62 
1.08 

- 

1.36 
.25 
1.15 

.88 

.96 

.87 
.94 
1.01 

Wave 3 

TRF 

Α 
- 

- 

- 

- 

- 

- 
- 

- 

- 
- 
- 

.76 

.38 

.63 
.43 
.58 

β 
- 

- 

- 

- 

- 

- 
- 

- 

- 
- 
- 

.60 

1.05 

- 

- 

YSR 

.30 
.46 
.53 

α 
.59 

.40 

.64 

.62 

.44 

.50 
.71 

.33 
-.08 
1.02 

Β 
.44 

1.38 

1.56 

.16 

1.09 

1.49 
1.67 

- 

- 

.28 
.65 
.40 

.56 
-.01 
1.66 

Note: *The CBCL item “disobedient at home” was used as the anchor item at each wave for 

model identification purposes. 

133 

 
 
 
 
 
 
 
 
 
 
 
Table S2.5. Standardized IRT parameter estimates from final model 

CBCL 

TRF 

YSR 

Wave 1 

Cruelty, 
bullying, or 
meanness 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 
Physically 
attacks 
people 
Explosive 
and 
unpredictable 
Hot temper 

Cruelty, 
bullying, or 
meanness 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 
Physically 
attacks 
people 
Explosive 
and 
unpredictable 
Hot temper 

α 
.36 

.55 

.46 

.37 

.72 

.35 
.33 

- 

β 
1.06 

.94 

.91 

.00 

.96 

1.39 
1.04 

Α 
.83 

.80 

.84 

- 

.93 

.86 
.91 

β 
1.14 

1.53 

1.72 

- 

1.00 

1.60 
1.60 

- 

.80 

1.54 

.25 

.27 

.79 
Wave 2 

1.44 

α 
- 

- 

- 

- 

- 

- 
- 

- 

- 

CBCL 

TRF 

YSR 

α 
.52 

.63 

.66 

.52 

.77 

.74 
.68 

- 

β 
1.00 

.80 

.65 

.00 

.47 

.97 
.92 

- 

.45 

.27 

Α 
.80 

.64 

.74 

- 

.86 

.92 
.88 

.96 

.93 

134 

β 
.23 

.98 

.94 

- 

.17 

.54 
.76 

.39 

.45 

β 
- 

- 

- 

- 

- 

- 
- 

- 

- 

Β 
.32 

.75 

1.09 

.20 

.62 

.94 
1.10 

α 
.44 

.55 

.65 

.33 

.59 

.60 
.57 

- 

- 

.44 

-.04 

 
 
 
 
 
Table S2.5 (cont’d) 

Wave 3 

CBCL 

TRF 

YSR 

Cruelty, 
bullying, or 
meanness 
Destroys 
own things 
Destroys 
things 
belonging to 
others 
Disobedient 
at home* 
Disobedient 
at school 
Many fights 
Physically 
attacks 
people 
Explosive 
and 
unpredictable 
Hot temper 

α 
.78 

.75 

.73 

.82 

.70 

.82 
.92 

- 

β 
.91 

1.10 

.96 

.00 

.86 

1.59 
1.06 

- 

.71 

.32 

Α 
- 

- 

- 

- 

- 

- 
- 

- 

- 

Β 
- 

- 

- 

- 

- 

- 
- 

- 

- 

α 
.59 

.42 

.72 

.50 

.57 

.44 
.64 

Β 
.43 

1.36 

1.46 

.28 

.95 

1.56 
1.75 

- 

- 

.67 

-.03 

Note: *The CBCL item “disobedient at home” was used as the anchor item at each wave for 

model identification purposes. 

135 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S2.6. LGM fit statistics and parameter estimates for multi-informant mean AGG scores 

Model 
Linear growth 
Means model 

Unconditional LGM model fit statistics 

-2lnL 
8049.46 
8189.34 

χ2 (df) 
- 
139.88† (3) 

AIC 
8061.46 
8195.34 

BIC 
8095.25 
8212.23 

SABIC 
8076.18 
8202.70 

Parameter 

Mean 
Variance 
ADI  
Income  
Conflict  
Nurturance  
Sex  
Race  

Mean 
Variance 
ADI 
Income  
Conflict  
Nurturance  
Sex  
Race  

Conditional LGM parameter estimates 
Estimate 

p-value 

.938 
.510 
.084 
-.094 
.335 
.001 
.317 
.032 

-.021 
.005 
-.003 
-.002 
-.018 
.003 
-.018 
-.007 

S.E. 
Intercept 
.076 
.046 
.031 
.034 
.026 
.024 
.053 
.079 
Slope 
.011 
.001 
.005 
.005 
.004 
.004 
.008 
.011 
Covariances 
.006 

<.001 
<.001 
.007 
.006 
<.001 
.977 
<.001 
.684 

.060 
<.001 
.567 
.698 
<.001 
.445 
.031 
.570 

-.036 

<.001 

.048 
.046 

.481 
.341 

Intercept with slope 
Intercept 1 with 
Intercept 2 
          MZ 
          DZ 
Slope 1 with Slope 2 
          MZ 
          DZ 
Intercept 1 with Slope 
2 
            MZ 
            DZ 
Residual variance 
Note: Bold font indicates p<.05. †Significant change in chi-square at p<.05.  

<.001 
<.001 
<.001 

-.040 
-.030 
.319 

.006 
.006 
.024 

<.001 
.001 

<.001 
<.001 

.001 
.001 

.006 
.003 

136 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S2.7. LCGA fit statistics and predictors of group membership for multi-informant mean 

AGG scores 

# of groups 
1 
2 
3 
4 
5 
6 

Unconditional LCGA model fit statistics 

AIC 

BIC 

Smallest group 
(% of sample) 

8292.06 
7893.68 
7797.62 
7713.73 
7676.40 
7639.27 

8308.95 
7927.46 
7848.29 
7781.29 
7760.85 
7740.61 

100 
45.6 
12.6 
3.3 
2.7 
2.9 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 

p-value 
<.001 
<.001 
<.001 
.216 
<.001 
.971 

Estimate 
.403 
-.393 
1.233 
.117 
.808 
.007 

Conditional LCGA parameter estimates from 3-group model 
Persistent v. low/stable 
S.E. 
.114 
.090 
.117 
.095 
.201 
.207 
Persistent v. desisting 
S.E. 
.162 
.134 
.139 
.134 
.308 
.317 
Desisting v. low/stable 
S.E. 
.104 
.123 
.106 
.098 
.195 
.257 

Estimate 
.054 
.056 
.881 
-.034 
.666 
.290 

Estimate 
.350 
-.449 
.352 
.151 
.142 
-.283 

p-value 
.607 
.646 
<.001 
.728 
.001 
.258 

p-value 
.031 
.001 
.011 
.258 
.645 
.372 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 
Note: Bold font indicates p<.05. 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 

137 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S2.8. Parameter estimates from conditional LGM with data in “long” format 

Parameter 

Estimate 

Mean 
Variance 
ADI  
Income  
Conflict  
Nurturance  
Sex  
Race  

Mean 
Variance 
ADI 
Income  
Conflict  
Nurturance  
Sex  
Race  

Intercept with 
slope 
Residual 
variance 

.708 
.286 
.054 
-.071 
.291 
.028 
.180 
.008 

-.047 
.002 
-.003 
.003 
-.019 
.000 
-.017 
.006 

-.020 

.149 

S.E. 
Intercept 
.047 
.025 
.019 
.021 
.017 
.017 
.033 
.048 
Slope 
.007 
.001 
.003 
.003 
.003 
.003 
.005 
.007 

.003 

.014 

p-value 

<.001 
<.001 
.005 
.001 
<.001 
.110 
<.001 
.870 

<.001 
<.001 
.394 
.380 
<.001 
.950 
.001 
.418 

<.001 

<.001 

Note: Parent-reported AGG is the outcome measure. Bold font indicates p<.05. 

138 

 
 
 
 
 
 
 
 
 
 
 
Table S2.9. LCGA fit statistics and predictors of AGG group membership with 1 twin randomly 

selected per family (N = 1,030) 

Unconditional LCGA model fit statistics 

AIC 

BIC 

Smallest group  
(% of sample) 

# of groups 
1 
2 
3 
4 
5 
6 

2976.49 
2708.76 
2589.30 
2533.19 
2474.60 
2454.29 

2991.28 
2738.35 
2633.69 
2592.38 
2548.59 
2543.07 

100 
38.7 
15.0 
9.4 
2.0 
2.7 

Conditional LCGA parameter estimates from 3-group model 
Persistent v. low/stable 
Estimate 
.281 
-.394 
1.221 
.158 
.684 
.320 

p-value 
.055 
.001 
<.001 
.313 
.010 
.347 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 

ADI 
Income 
Conflict 
Nurturance 
Sex 
Race 
Note: Parent-reported AGG is the outcome measure. Bold font indicates p<.05.  

S.E. 
.146 
.120 
.161 
.156 
.265 
.340 
Persistent v. desisting 
S.E. 
Estimate 
.189 
.180 
.135 
-.358 
.166 
.467 
.201 
.169 
.335 
.169 
.418 
.167 
Desisting v. low/stable 
S.E. 
Estimate 
.120 
.101 
.114 
-.036 
.115 
.755 
.117 
-.011 
.204 
.514 
.271 
.153 

139 

p-value 
.340 
.008 
.005 
.401 
.613 
.689 

p-value 
.400 
.753 
<.001 
.924 
.012 
.573 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Figures 

Males

Females

G
G
A

1

0.8

0.6

0.4

0.2

0

6

7

8

9

10 11 12 13 14 15 16 17 18 19 20 21

Age

Figure S2.1. Mean patterns of AGG for males and females from the latent growth curve model 

(variable-centered).  

140 

 
 
  
 
G
G
A

2

1.6

1.2

0.8

0.4

0

Group 1: low, stable
(52.1%)

Group 2: desisting
(12.6%)

Group 3: persisting
(35.4%)

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 21

Age

Figure S2.2. Unconditional three-group model of AGG trajectories based on multi-informant 

mean scores. 

141 

 
 
 
 
APPENDIX C: SUPPLEMENTAL MATERIAL FOR CHAPTER 3 

Supplementary Methods 

Data Analytic Strategy 

Item Response Theory (IRT) Analyses  

Analyses were conducted in Mplus 8.4 (Muthén & Muthén, 1998-2019) using the 

weighted least squares mean- and variance-adjusted (WLSMV) estimator. As the twins were 

nested within families, analyses accounted for the nonindependence of observations using the 

CLUSTER command. We made use of an IRT model, which relates an individual’s trait level (θ) 

to their performance on a series of items while accounting for item characteristics. Items were 

drawn from the Rule-Breaking (RB) scale on the ASEBA instruments, including the Child 

Behavior Checklist (CBCL), Teacher Report Form (TRF), and Youth Self-Report (YSR) 

(Achenbach & Rescorla, 2001). Given the relatively low frequency of RB in the sample, and 

consistent with prior IRT analyses of antisocial behavior in the TBED-C (Burt, Donnellan, 

Slawinski, & Klump, 2016), items were coded dichotomously, with responses of 1 or 2 on the 

ASEBA instruments collapsed into a single “present” category (i.e., 0=behavior absent, 

1=behavior present). We then applied the two-parameter logistic (2PL) model, which estimates 

two item characteristics: discrimination (α) and difficulty (β) (Embretson & Reise, 2013). Item 

discrimination indicates the extent to which endorsement of a given item relates to one’s trait 

level; the higher the discrimination, the more accurately the item assesses standing on the trait 

of interest. Item difficulty indicates the trait level needed to have a 50% chance of endorsing the 

item. A higher difficulty thus indicates that one would need a relatively high standing on the 

latent trait continuum to receive a 1. Such an item would be diagnostic of high levels of RB. 

We omitted items assessing substance use and sexual problems, given the very low 

base rates of these behaviors in childhood. At Wave 1, for example, no informant indicated that 

drug use was a concern, and only one teacher endorsed the item “seems preoccupied with 

sex”. Other behaviors that were endorsed at low rates during Wave 1 were not endorsed at all 

142 

 
during later assessments; for instance, no caregiver endorsed “sets fires” at Wave 3 or 

“vandalism” at Waves 2 or 3. In addition, the item “runs away from home” was endorsed by less 

than 2% of caregivers and youth at Waves 2 and 3 (this item is not included on the TRF). Thus, 

initial models included 18 items at Waves 1 and 3 and 27 items at Wave 2 (i.e., 9 items from 

each available informant) assessing behaviors that were present to some degree at all 

assessments. Each wave comprised one latent factor. Items were omitted from the final model 

based on their characteristics (i.e., low discrimination and/or high difficulty). Model fit was 

evaluated with three indices: root mean square error of approximation (RMSEA), Comparative 

Fit Index (CFI), and Tucker-Lewis Index (TLI). RMSEA values below .06 and CFI and TLI values 

above .95 are considered to indicate good fit (Hu & Bentler, 1999). For each model, we 

examined modification indices to identify sources of misfit and added the suggested residual 

covariances between items to the model. RB mean scores were computed based on 

participants’ responses to the items retained in the final IRT model. 

Measurement Invariance Analyses 

After selecting the items that functioned well at each assessment wave, we conducted a 

series of analyses to test for measurement invariance (i.e., the extent to which items assessed 

the same construct over time, a prerequisite for examining developmental change; Widaman, 

Ferrer, & Conger, 2010). Analyses made use of the weighted least squares mean and variance 

adjusted (WLSMV) estimator and delta parameterization in Mplus 8.4, an appropriate method 

for dichotomous outcome data (Múthen & Asparouhov, 2002). The initial model examined 

configural invariance, with all items from a given wave loading onto a single factor and item 

loadings and thresholds free to vary across waves. Factor means and variances were fixed to 0 

and 1, respectively, and scale factors were fixed to 1, for model identification purposes. 

Modification indices were examined to identify sources of misfit, and the suggested residual 

covariances between items were added to the model. The second model tested for scalar 

invariance, with item loadings and thresholds constrained to equality within each informant over 

143 

 
time. To identify the model, the mean and variance of the latent factor at Wave 1 were 

constrained to 0 and 1, respectively, and the scale factors were fixed to 1 for each informant at 

one time point (i.e., scale factors for CBCL and TRF items were fixed to 1 at Wave 1, and scale 

factors for YSR items were fixed to 1 at Wave 2). The same residual covariances between items 

that were included in the configural invariance model were also included when assessing scalar 

invariance. Model fit was assessed using the RMSEA, CFI, TLI, and χ2 difference test (Múthen, 

du Toit, & Spisic, 1997). Minimal changes in fit between the two models would indicate that the 

items measured RB consistently across assessment waves. (Other models to assess 

measurement invariance, such as a metric invariant model that allows item thresholds, but not 

loadings, to vary across waves, are not identified when examining binary items (Múthen & 

Asparouhov, 2002). As a result, we focused on the respective fits of the configural and scalar 

invariance models.)   

IRT Modeling 

Supplementary Results 

The initial IRT model at each wave included 9 ASEBA RB items from each available 

informant. We subsequently omitted the CBCL/TRF item “prefers being with older kids” and 

corresponding YSR item “prefer being with younger kids” due to their low discrimination values. 

Fit indices are reported in Table S3.2, and parameter estimates are reported in Tables S3.3 and 

S3.4. Fit was good at Wave 1 according to all fit indices. At Waves 2 and 3, the CFI and TLI 

indicated some degree of misfit, even after residual covariances were added, but the RMSEA 

was well below recommended cutoffs (Hooper, Coughlan, & Mullen, 2008; Hu & Bentler, 1999; 

Steiger, 2007), indicating good fit. The final model at each wave retained eight items from each 

available informant. As expected, items assessing illegal behaviors (e.g., steals) were more 

difficult to endorse than were those assessing relatively minor rule violations (e.g., lying or 

cheating). Some items (e.g., truancy) were much easier to endorse at later assessment waves 

relative to Wave 1, reflecting the greater frequency of those behaviors during adolescence. 

144 

 
Teacher-reported items generally discriminated best, whereas self-reported items had the 

lowest discrimination values. We computed mean RB scores based on the items retained in the 

final model. Scores were multiplied by 24 to correspond to the highest number of items included 

at any wave (i.e., Wave 2) and then log-transformed to account for positive skew. 

Measurement Invariance 

Model fit estimates from the measurement invariance analyses are shown in Table S3.5. 

The χ2 difference test was significant, suggesting decrement in model fit. However, the χ2 fit 

statistic is heavily influenced by sample size (Bentler & Bonett, 1980) (see Yuan & Chan (2016) 

for a discussion of limitations specific to χ2 difference testing). Moreover, the absolute fit indices 

were nearly identical for the two models, meaning that both the item loadings and thresholds 

could be constrained to equality across waves (Putnick & Bornstein, 2016). The items thus 

appeared to assess RB in a way that was consistent and developmentally appropriate across 

assessment waves. 

145 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.1. Youth RB by age 

Tables 

Age 
6 
7 
8 
9 
10 
11 
12 
13 
14 
15 
16 
17 
18 
19 
20-21 

Total N 
599 
414 
347 
349 
354 
119 
82 
92 
186 
234 
170 
136 
68 
36 
7 

Mean RB (SD) 
2.80 (3.45) 
2.48 (3.27) 
2.74 (3.39) 
2.14 (3.48) 
2.42 (3.44) 
2.31 (2.95) 
2.14 (2.19) 
2.48 (2.89) 
2.67 (3.43) 
2.72 (3.21) 
2.87 (3.17) 
3.69 (3.47) 
2.81 (3.13) 
3.24 (4.01) 
4.50 (3.35) 

Range RB 
0-21 
0-18 
0-21 
0-21 
0-21 
0-15 
0-10 
0-12 
0-23 
0-15 
0-21 
0-15 
0-16.5 
0-18 
0-9 

Note: Multi-informant composite mean RB scores are the outcome measure. 

146 

 
 
 
Table S3.2. IRT model fit 

Wave 1 

Wave 2 

Wave 3 

Model 1 
Model 2 

Model 1 
Model 2 

Model 1 
Model 2 

RMSEA (90% CI) 

.019 (.015-.024) 
.021 (.016-.026) 

.028 (.024-.033) 
.032 (.028-.037) 

.040 (.029-.050) 
.044 (.032-.055) 

CFI 

.981 
.984 

.935 
.936 

.917 
.923 

TLI 

.973 
.975 

.925 
.924 

.900 
.905 

Note: Model 1 contained 9 ASEBA RB items from each available informant at a given wave. 

Model 2 was the final model, containing items that were endorsed at all waves and had good 

psychometric properties.  

147 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.3. Standardized IRT parameter estimates from Model 1 at each wave 

Wave 1 

CBCL 

TRF 

Breaks rules* 
Doesn’t seem to feel 
guilty after misbehaving 
Hangs around with 
others who get in trouble 
Lies or cheats 
Prefers being with older 
kids 
Prefers being with 
younger kids 
Steals at home 
Steals outside the home 
Steals 
Swears 
Truant 
Tardy to school or class 

Breaks rules* 
Doesn’t seem to feel 
guilty after misbehaving 
Hangs around with 
others who get in trouble 
Lies or cheats 
Prefers being with older 
kids 
Prefers being with 
younger kids 
Steals at home 
Steals outside the home 
Steals 
Swears 
Truant 
Tardy to school or class 

Breaks rules* 
Doesn’t seem to feel 
guilty after misbehaving 

α 
.59 
.51 

.50 

.60 
.18 

- 

.63 
.65 
- 
.61 
.54 
- 

α 
.62 
.54 

.69 

.51 
.36 

- 

.73 
.87 
- 
.47 
.88 
- 

α 
.76 
.66 

β 
.00 
.69 

.99 

.31 
.35 

- 

1.25 
1.53 
- 
1.06 
2.24 
- 
Wave 2 

CBCL 

β 
.00 
.16 

.30 

.25 
.30 

- 

.62 
.70 
- 
.15 
.68 
- 
Wave 3 

CBCL 

β 
.00 
.19 

148 

α 
.89 
.75 

.72 

.82 
.27 

- 

- 
- 
.75 
.84 
.29 
.34 

α 
.94 
.93 

.81 

.91 
.29 

- 

- 
- 
.78 
.79 
.67 
.66 

α 
- 
- 

Β 
.64 
.87 

.80 

.96 
1.48 

- 

- 
- 
1.59 
1.76 
1.86 
1.19 

Β 
-.30 
-.35 

-.32 

-.02 
.85 

- 

- 
- 
.77 
.08 
.21 
-.07 

Β 
- 
- 

TRF 

TRF 

YSR 
α 
- 
- 

- 

- 

- 

- 
- 
- 
- 
- 
- 

β 
- 
- 

- 

- 

- 

- 
- 
- 
- 
- 
- 

YSR 
α 
.35 
.37 

β 
.28 
.27 

.39 

.36 

.19 
- 

.81 
- 

-.02 

.81 

.51 
.60 
- 
.23 
.63 
- 

.97 
1.21 
- 
-.30 
.72 
- 

YSR 
α 
.63 
.34 

β 
.16 
.54 

 
 
 
 
 
 
 
 
 
 
 
 
Table S3.3 (cont’d) 

- 

- 

- 
- 

.34 

.49 

.83 

.32 

.69 
.31 

-.06 
.53 

Hangs around with 
others who get in trouble 
Lies or cheats 
Prefers being with older 
kids 
Prefers being with 
younger kids 
1.26 
.91 
Steals at home 
.88 
.88 
Steals outside the home 
- 
- 
Steals 
-.89 
.09 
Swears 
.62 
.63 
Truant 
Tardy to school or class 
- 
- 
Note: *The CBCL item “breaks rules” was used as the anchor item at each wave for model 

.95 
.91 
- 
.57 
.76 
- 

.50 
.76 
- 
.58 
.54 
- 

- 
- 
- 
- 
- 
- 

- 
- 
- 
- 
- 
- 

.18 
- 

.90 
- 

-.47 

.37 

- 
- 

- 

- 

- 

- 

identification purposes. 

149 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.4. Standardized IRT parameter estimates from Model 2 at each wave 

Breaks rules* 
Doesn’t seem to feel 
guilty after 
misbehaving 
Hangs around with 
others who get in 
trouble 
Lies or cheats 
Steals at home 
Steals outside the 
home 
Steals 
Swears 
Truant 
Tardy to school or 
class 

Breaks rules* 
Doesn’t seem to feel 
guilty after 
misbehaving 
Hangs around with 
others who get in 
trouble 
Lies or cheats 
Steals at home 
Steals outside the 
home 
Steals 
Swears 
Truant 
Tardy to school or 
class 

Breaks rules* 
Doesn’t seem to feel 
guilty after 
misbehaving 

Wave 1 

CBCL 

TRF 

β 
.00 
.68 

α 
.88 
.84 

β 
.63 
.80 

YSR 
α 
- 
- 

β 
- 
- 

.98 

.71 

.80 

.31 
1.24 
1.50 

- 
1.06 
2.23 
- 

Wave 2 

CBCL 

β 
.00 
.20 

.85 
- 
- 

.75 
.83 
.29 
.34 

α 
.94 
.93 

.92 
- 
- 

1.58 
1.76 
1.85 
1.19 

β 
-.30 
-.35 

TRF 

- 

- 
- 

- 
- 
- 
- 

- 

- 
- 

- 
- 
- 
- 

YSR 
α 
.36 
.39 

β 
.27 
.25 

.30 

.81 

-.31 

.39 

.35 

.26 
.63 
.70 

- 
.15 
.67 
- 

.91 
- 
- 

.78 
.78 
.68 
.66 

Wave 3 

CBCL 

TRF 

β 
.00 
.17 

α 
- 
- 

-.01 
- 
- 

.77 
.09 
.21 
-.07 

β 
- 
- 

.20 
.52 
.61 

- 
.22 
.64 
- 

.80 
.96 
1.19 

- 
-.29 
.71 
- 

YSR 
α 
.61 
.35 

β 
.19 
.53 

α 
.57 
.50 

.50 

.59 
.64 
.67 

- 
.60 
.54 
- 

α 
.62 
.53 

.68 

.51 
.72 
.87 

- 
.47 
.89 
- 

α 
.77 
.67 

150 

 
 
 
 
 
 
 
 
 
 
 
Table S3.4 (cont’d) 

Hangs around with 
others who get in 
trouble 
Lies or cheats 
Steals at home 
Steals outside the 
home 
Steals 
Swears 
Truant 
Tardy to school or 
class 

.82 

.71 
.94 
.90 

- 
.57 
.77 
- 

.35 

-.07 
.93 
.91 

- 
.11 
.64 
- 

- 

- 
- 
- 

- 
- 
- 
- 

- 

- 
- 
- 

- 
- 
- 
- 

.50 

.33 

.20 
.49 
.76 

- 
.57 
.55 
- 

.86 
1.28 
.89 

- 
-.87 
.62 
- 

Note: *The CBCL item “breaks rules” was used as the anchor item at each wave for model 

identification purposes. 

151 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.5. Measurement invariance results 

χ2 difference test 
(df) 
- 

RMSEA (90% CI) 

.012 (.010-.014) 

51.75† (28) 

.012 (.010-.013) 

CFI 

.920 

.920 

TLI 

.915 

.916 

Configural 
Invariance 
Scalar 
Invariance 

Note: †Significant change in chi-square at p<.05.  

152 

 
  
 
 
 
Table S3.6. Predictors of LCGA group membership for remaining pairings 

Conditional LCGA parameter estimates from 4-group model 
Chronic v. escalating 

Estimate 
-.007 
-.475 
.311 

S.E. 
.132 
.121 
.189 

1.144 
.056 

.266 
.278 
Escalating v. desisting 
S.E. 
.141 
.123 
.218 

Estimate 
.185 
.119 
.902 

-.563 
.034 

.255 
.313 
Desisting v. normative 
S.E. 
.088 
.093 
.154 

Estimate 
.275 
-.278 
.860 

p-value 
.955 
<.001 
.101 

<.001 
.840 

p-value 
.190 
.336 
<.001 

.028 
.913 

p-value 
.002 
.003 
<.001 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

.151 
.224 
Note: Multi-informant composite mean RB scores are the outcome measure. Bold font indicates 

.058 
.291 

.286 
.237 

p< .05. 

153 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.7. LGM fit statistics and parameter estimates for caregiver reports of RB 

Model 
Linear growth 
Means model 

Unconditional LGM model fit statistics 

-2lnL 
5101.84 
5205.66 

χ2 (df) 
- 

AIC 
5113.84 
103.82† (3)  5211.66 

BIC 
5147.62 
5228.55 

SABIC 
5128.56 
5219.02 

Parameter 

Mean 
Variance 
ADI  
Income  
Peer Delinquency  
Sex  
Race  

Mean 
Variance 
ADI 
Income  
Peer Delinquency  
Sex  
Race  

Intercept with 
slope 
Intercept 1 with 
Intercept 2 
          MZ 
          DZ 
Slope 1 with Slope 2 
          MZ 
          DZ 
Intercept 1 with 
Slope 2 
            MZ 
            DZ 
Residual variance 

Conditional LGM parameter estimates 
Estimate 

p-value 

S.E. 
Intercept 
.049 
.018 
.020 
.021 
.018 
.034 
.052 
Slope 
.007 
.000 
.003 
.004 
.004 
.006 
.008 
Covariances 
.002 

.017 
.018 

.000 
.000 

.002 
.002 
.010 

<.001 
<.001 
.141 
.002 
<.001 
<.001 
.167 

.150 
<.001 
.868 
.925 
.102 
<.001 
.950 

<.001 

<.001 
<.001 

<.001 
<.001 

<.001 
<.001 
<.001 

.410 
.225 
.029 
-.067 
.134 
.171 
.071 

-.011 
.003 
-.001 
.000 
.006 
-.021 
.000 

-.017 

.245 
.154 

.003 
.001 

-.021 
-.012 
.120 

Note: RB sum scores were computed based on maternal reports on eight CBCL items. Bold font 

indicates p<.05. †Significant change in chi-square at p<.05.  

154 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.8. LCGA fit statistics and predictors of group membership for caregiver reports of RB 

Unconditional LCGA model fit statistics 

AIC 

BIC 

Smallest group 
(% of sample) 

# of groups 
1 
2 
3 
4 
5 
6 

5363.62 
4857.79 
4637.80 
4461.54 
4345.10 
4239.47 

5380.50 
4891.54 
4688.43 
4529.04 
4429.48 
4340.73 

100 
24.9 
14.5 
3.3 
2.5 
2.3 

Conditional LCGA parameter estimates from 4-group model 
Chronic v. normative 
Estimate 
.166 
-.400 
.954 

S.E. 
.101 
.089 
.118 

p-value 
.102 
<.001 
<.001 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

.824 
.443 

.221 
.238 
Escalating v. normative 
S.E. 
Estimate 
.186 
.052 
.144 
-.210 
.155 
1.020 

-.220 
.515 
Chronic v. desisting 

.341 
.338 

<.001 
.063 

p-value 
.780 
.145 
<.001 

.519 
.128 

Estimate 
-.018 
-.289 
.650 

S.E. 
.123 
.115 
.118 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 
Note: RB sum scores were computed based on maternal reports on eight CBCL items. Bold font 

.254 
.290 

.071 
.888 

.459 
.041 

p-value 
.884 
.012 
<.001 

indicates p<.05. 

155 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.9. Parameter estimates from conditional LGM with data in “long” format 

Parameter 

Mean 
Variance 
ADI  
Income  
Peer Delinquency  
Sex  
Race  

Mean 
Variance 
ADI 
Income  
Peer Delinquency  
Sex  
Race  

Intercept with 
slope 
Residual variance 

Estimate 
Intercept 
.651 
.442 
.055 
-.159 
.183 
.250 
.061 
Slope 
.030 
.005 
.002 
.012 
.021 
-.023 
-.002 

-.037 

S.E. 

.061 
.042 
.025 
.027 
.026 
.046 
.063 

.010 
.001 
.004 
.005 
.004 
.007 
.010 

.005 

p-value 

<.001 
<.001 
.028 
<.001 
<.001 
<.001 
.333 

.002 
<.001 
.709 
.007 
<.001 
.002 
.869 

<.001 

.023 
Note: Multi-informant composite mean RB scores are the outcome measure. Bold font indicates 

<.001 

.326 

p<.05. 

156 

 
 
 
 
 
 
 
 
 
 
 
 
 
Table S3.10. LCGA fit statistics and predictors of RB group membership with 1 twin randomly 

selected per family (N = 1,030) 

Unconditional LCGA model fit statistics 

AIC 

BIC 

Smallest group  
(% of sample) 

# of groups 
1 
2 
3 
4 
5 
6 

3997.34 
3833.60 
3774.99 
3693.81 
3653.97 
3624.97 

4012.15 
3863.22 
3819.41 
3753.03 
3728.00 
3713.81 

100 
39.4 
5.0 
8.1 
4.2 
2.8 

Conditional LCGA parameter estimates from 4-group model 
Chronic v. normative 
S.E. 
Estimate 
.200 
.573 
.163 
-.384 
.283 
1.939 

p-value 
.004 
.018 
<.001 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

ADI 
Income 
Peer 
Delinquency 
Sex 
Race 

.402 
.327 

.317 
.378 

Escalating v. normative 
Estimate 
.705 
.109 
1.671 

S.E. 
.207 
.219 
.280 

-.291 
.364 
.489 
.351 
Chronic v. desisting 
S.E. 
.235 
.166 
.262 

Estimate 
.211 
-.169 
1.246 

.102 
.066 

.371 
.406 

.205 
.387 

p-value 
.001 
.618 
<.001 

.423 
.474 

p-value 
.369 
.309 
<.001 

.784 
.870 

Note: Multi-informant composite mean RB scores are the outcome measure. Bold font indicates 

p<.05.  

157 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
i

g
n
k
a
e
r
B
-
e
u
R

l

Figures 

2.5

2

1.5

1

0.5

0

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Age

Group 1: normative
(50.10%)

Group 2: escalating
(3.27%)

Group 3:
moderate/declining
(33.02%)

Group 4:
high/declining
(13.61%)

Figure S3.1. Unconditional four-group model of RB trajectories based on maternal reports at all 

waves. 

158